Show IH: I built 18 industry pages for the small businesses AI search ignores — which one did I leave out?

Thanks. The starting list came from two simple filters. First, high value per customer, so trades and professional services where one missed recommendation is worth real money, like HVAC, lawyers, med spas, dentists. Second, industries where buyers ask AI in full sentences ("who do I call for X near me") rather than by name, because that is exactly the query shape that produces a short three name answer and leaves everyone else out.

Basically I looked for the overlap of "expensive to be invisible" and "buyers already asking AI this way." That is where the gap hurts most, so it felt like the right place to begin.

Thanks, and I like your angle a lot. Localized AI music for French and Japanese is the same bet in a different field, going specific where the big global tools stay generic and miss the nuance. That is exactly where independents win.

On update frequency, the honest answer is the landing pages themselves do not need frequent updates, because they are thin on purpose. The real content lives in the report, which is generated fresh per business each time from live test results, so it is always current without me editing pages. What I do watch is the buyer question set per industry, since the way people ask shifts over time and the models change. So the pages stay fairly stable, and my update effort goes into keeping the questions and the test logic current rather than rewriting 18 pages constantly. Good luck with the music platforms.

Great add, and the rationale is exactly right. Independent B2B consultants and fractional operators are a strong fit because their buyers ask very specific, high intent questions like "fractional ops lead for a manufacturing company," yet the best practitioners often have thin sites and live on referrals. That is the perfect setup for the gap this catches, known by name, invisible on the actual buyer problem. Someone else in the thread flagged them too, so they are going in.

And you put it better than I have, the discoverable by name versus discoverable for a buyer problem split is the clearest way to make the value tangible for this group, since their whole reputation lives outside the site.

Glad the "one useful improvement with an effort estimate" format lands too. That is the point of it, turning visibility from an abstract SEO talk into a decision an owner can actually make this week.

Nice, and I think we landed on the same insight from opposite ends. You went deep on one niche with 34 specific articles, I went wide across 18 industries, but both of us are really saying the same thing. Broad "what is X" queries are dead because the AI overview just answers them itself. The value is in the hyper specific question the summary cannot cover.

Your sourdough example is a perfect proof of it. "65 hydration sourdough" beating "how to bake sourdough" even with fewer backlinks is exactly the pattern. The person asking the specific thing wants the exact number, not the Wikipedia summary, so specificity wins over authority there. That is the same reason a business gets recommended for a precise buyer question but not the generic one.

And going from zero to 79 weekly impressions in 4 months from real intent people is a solid result. Slow but real beats big and empty. Nice work.

The problem you are describing for local businesses is exactly what happens to SaaS founders who vibe-coded their MVP. The product exists, the model knows the name, but when a buyer asks "what tool helps with X" the answer is always the three competitors with real documentation behind them. The vibe-coded founder has the same gap as the HVAC contractor, citable proof is missing, not the product.

The difference is the founder usually does not know it either. No alert, no signal. Just quieter than expected inbound.

Software development firms and custom build teams would be an interesting industry to add. Buyers ask "how do I find a technical team to take over my product" and almost always get generic platforms back, never the right small team

Fully agree, and this is the point a few people have pushed me on, rightly. One snapshot is a hook, not truth. AI answers shift day to day and even hour to hour, so running the same eight questions three times in one day to get a baseline variance before recommending anything is exactly the right move. Without that band, a fix looks like it worked when the model just wandered, or looks dead when it actually helped. That is the next build, and pairing it with a frozen control so a change is attributable to the edit and not a quiet model update.

On the most surprising gap, honestly the high ticket trades like HVAC and roofing. You would expect industries with that much money moving through them to be all over this, but the sites are often just a phone number and a photo gallery. So the model knows the name fine and then names five competitors on the real buyer question, because there is nothing citable to work with. Big money, wide open, that surprised me most.

Agreed on all of this, and it is the direction I am moving. Running the same eight questions three times on the same day, then again after each change, is the clean way to do it. Reporting the range instead of just the best count is a small change that makes a big difference for trust, since an owner can see the difference between a real gain and the model simply having a different day.

Your simple log idea is exactly right too. Question, date, answer, and whether the business showed up in the top three. Kept over time and across businesses, that stops being a one off report and becomes a real benchmark. A few people in this thread have pushed me toward the same thing, so it is clearly the next build, not a nice to have.

Thanks, and you are right, structured content is a big part of why smaller companies disappear. The engines lean on clear, consistent, structured info, and small sites often do not have it.

On the pages, it is a shared framework adapted per niche, not fully manual each time. The structure stays the same across all of them, but the pieces that matter, the buyer questions, the example gap, the competitor language, are built from real data for each specific industry. So the frame is shared, the content inside is per niche. I keep an eye on that so they do not drift into one generic template with the industry name swapped, since the whole value is in the industry specific part.

Fair point, and you are right that "here is your starting point" alone is not enough to be worth much. A score people read once and forget has little value. So the report is built to be actionable, not just a diagnosis. It does not say "improve your visibility," it says "add this homepage block, it takes an hour" and "publish an FAQ covering these exact questions buyers asked." Concrete fixes with rough time estimates, ordered fastest first.

Also worth saying, the free report is only the front door. If you log in, planmoon does the full end to end. It connects to your site and Search Console, runs the analysis, then gives you a content plan and strategy tailored to your site for blog, Instagram and LinkedIn. You can generate the actual content from that plan, and there is a WordPress plugin that publishes it automatically. So the goal is not just to tell you the gap, it is to close it for you.

Totally fair that you have not seen this yet since your site is password protected. But actionable advice is exactly the main selling point you were hoping for, so I would love your read once you can run it.

One honest note: it is a free limited beta right now, so we are still testing and smoothing some rough edges. Feedback like yours is exactly what helps.

Thanks, and you are pointing at the right thing. Honest answer is I am not really measuring quality well yet. Right now it is mostly early signals like which industry pages get URLs dropped and which get people asking for a report. That tells me interest, not real quality.

Your point about what happens after the first traffic is the part I care about most and have the least data on. For this tool the real quality signal is not a visit, it is whether an owner actually goes and ships the fixes, then comes back to check if the result moved. That loop is exactly what I want to build next, because a visibility score people read once and forget is worth very little. Which workflows people return to is the thing that will tell me if this is actually useful or just interesting.

Thanks , glad the "2 of 8" framing landed for you.

On healthcare and therapy clinics, that one is partly covered already. There is a Health and wellness section on planmoon.app/for, so that side of your suggestion is live and you can see how the framing works there.

Commercial construction and fit-out contractors is a strong new add though. High value, high consideration, and buyers really do ask "how do I choose a contractor for X" and end up seeing competitors. Adding it to the next batch.

And yes, the cross-industry dataset is the part I am most curious about too.

Thanks, and I agree with you completely. The technology is rarely the hard part, understanding what users actually need is. Same lesson on my side. The tool only works because it starts from the real questions buyers ask, not from what is easy to build. Good luck with your agent projects.

You are hitting the point I think planmoon fails to communicate. You are given actionable ideas on how to make AI knowledgebases include your site or product.

This matches what I am seeing exactly. The engines reward the same name, same description, same specifics repeated on the business's own site and on the directories it is already cited on. Owners usually fix one side and ignore the other, then wonder why nothing moved.

Your consulting point is spot on too. Vague services language like "strategic advisory" gives the model nothing to match to a real buyer question, while "discovery call to proposal for B2B SaaS" is specific enough to grab onto. So on-page specificity is the easier win, and it is where I start, because the owner fully controls it and it takes an hour.

But you are right that for solo and small service businesses the off-site consensus is the harder and often bigger lever. A dentist has software pushing consistent listings everywhere. A solo consultant does not, so their evidence stays scattered across LinkedIn, a directory or two, and a few talks, all describing them differently. My honest read is that on-page gets you understood, off-site gets you recommended, and the second one is the real work for small service businesses. That is the part I am still figuring out how to make as fast and simple as the homepage fix.

Thanks. Logistics and courier services is a good add, lots of "who can ship X" and "same day courier near me" type questions, so adding it.

On measurable improvement, honest answer is not yet. The pages are new and this version has not been in front of many people, so I do not have solid before and after data. That is actually the next thing I want to build, rerunning the same questions after a business makes the fixes to see what really moves. Right now it is one snapshot per business, not a tracked change over time.

Thanks. Your angle is interesting, and yes I run into that gap a lot. Most owners I talk to know sales in a rough way but have not turned it into a "focus here or not" decision, exactly like you said. It is usually Excel and gut feeling, not numbers they act on.

For my side it shows up a bit differently. The visibility fixes are cheap enough that an owner does not really need to know their numbers to justify them, an hour of writing is easy to say yes to. But knowing the numbers changes which fixes matter. An owner who knows a single client is worth years of revenue treats a missing recommendation very differently from one who does not. So the ones who know their numbers act faster and pick better, and the ones who do not tend to feel the problem but not move on it. Sounds like your tool is aimed right at that second group.

Yes, that is the whole game. The branded query is just a vanity check, the buyer query is the one that actually spends money. I like your three bucket idea a lot: you, a named competitor, or nobody. One row per question is a clean way to log it, and running fresh chats matters because the shortlist really does jitter between runs, so a single pass can fool you.

Your last point matches what I see too. If the model cannot tell who you help from the first screen, it just names five other people instead. That is exactly why the homepage block goes first in the fixes. Fastest way to give the model something clear to work with.

Accountants are already live, actually. planmoon.app/for/accountants

You picked a good one. They are very local and trust driven, and people really do ask "best accountant for small businesses" or "who can help with X tax issue." That is exactly the kind of full sentence buyer question that produces a short three name answer and leaves everyone else out.

Seeing how often AI recommends them versus traditional search is the interesting part, and it is basically what the report checks.

Thanks, and agreed. Generic advice misses a lot because the buyer questions really are different per industry.

Auto repair is a great pick, and a few other people said the same in this thread. Very high intent, very local, lots of "who can fix this today" type questions. Adding it.

Thanks! Cross-border e-commerce and B2B trade services are both interesting adds , higher-consideration, lots of "how do I find a supplier/partner who does X" questions, which is the query shape this catches well. Adding them to the list.

Thanks! Biggest visibility gaps so far are the high-ticket trades — HVAC, roofing, med spas. Counterintuitive, since you'd expect industries with that much money moving through them to have someone minding the AI channel, but the sites are often just a phone number and a photo gallery, so there's nothing citable for the buyer question. The boring, expensive verticals turned out to be the most wide open.

Thanks. Tutoring and test prep is a great add. Parents really do ask AI "best SAT tutor near me" or "IB math tutor online," and like dentists or med spas it is very local plus niche credibility. I have not seen anyone build AI visibility tooling for it either. Adding it.

On your scoring question, the questions are customized per industry, not one fixed set of 8 templates. You are right that the way buyers phrase things for an immigration lawyer is very different from an HVAC contractor, so reusing the same wording would give bad results. For each industry I build the 8 questions from real buyer language for that trade, then keep the ones where the AI has to actually know the business to answer well. The structure stays the same across industries, but the actual questions do not.

Agreed on the core — being understandable to the AI the same way you'd be to a human buyer is exactly it; content that doesn't answer the actual question loses even from a strong site.

The parallel holds too: visible-but-not-relevant-at-the-moment is the same failure whether it's a business missing the buyer question or a post missing the opening. Different layer, same gap. Good luck with HookFlow.

That citation list is the product, honestly. If the model cited matomo.org, posthog.com, plausible.io and indiehackers.com for one business, then those four domains are that business's actual target list, and you already generate it as a by-product of running the report.

The reason aggregators win is structural rather than unfair. They publish the sentence a model wants to lift: entity, category, location, one differentiator, in a consistent shape, repeated across thousands of records. Most small business homepages open with a mood instead of a definition, so there is nothing quotable to extract even when the page does get read.

Which suggests the on-page work is not dead, just much narrower than an industry landing page. One paragraph stating plainly what the business is, where it operates and what it does differently, phrased so it can be quoted without editing. Cheap to add, and it gives the model something to agree with when it goes looking for consensus.

Do your reports show the same aggregators repeating across unrelated verticals, or does each industry have its own set?

You've answered your own last question with the point that precedes it, and the honest data backs you up: in the reports I've run, it's mostly the aggregators. The Amami test leaned on matomo.org, posthog.com, plausible.io, indiehackers.com — competitor sites and directories, barely the business's own domain. ZEAM leaned on Instagram, Behance, Reddit, and a competitor's site. So yes — the model is largely repeating third-party consensus, not reading the business's homepage and being persuaded.

Which means your central point is one I have to concede against my own product: for a lot of these businesses, another page on their own site is not the highest-leverage move — getting one consistent description of what they do and where onto sources that aren't theirs is. My reports already nudge at this (the "directory profiles where competitors are already cited" recommendation), but you're framing it more sharply than I have: the owned page makes you citable, the off-site consensus makes you named. Both matter, but if I'm honest about where the leverage sits for a solo dentist, it's often the second one — and that's harder to sell as a one-hour fix, which is probably why I've under-weighted it.

The entity disambiguation trap is a real one I hadn't named explicitly — when the business name collides with a more famous meaning of the word, the model resolves toward the famous one and the business loses before content even enters the picture. That's a distinct failure mode from "thin content," and it needs a distinct fix (consistent entity signals, sameAs links, unambiguous name-plus-descriptor pairing). Worth flagging separately in reports rather than lumping into "add more content."

Opticians and audiologists is a great add for exactly the reason you said — insurance-attached, high-consideration, and asked in full sentences, which is the query shape that produces a three-name answer. Adding both.

This is the second or third comment converging on the same conclusion — that the evidence chain and off-site consensus matter as much as on-page content — so it's clearly the thing I need to build toward next rather than treat as an edge case.

Early read: the mechanic is the same across industries — known by name, missing on the buyer question — but the evidence that fixes it differs a lot. Local services win on directory consensus and reviews; SaaS wins on case studies and comparison pages. Same disease, different medicine.

Dentists are already live, actually — planmoon.app/for/dentists — along with veterinary clinics (planmoon.app/for/veterinary-clinics), so you can see how the framing shifts between two local-but-different categories.

Auto repair is a great pick — urgent, hyper-local, "who can fix this today," exactly the query shape that produces a three-name answer and leaves everyone else out of the consideration set entirely. Adding it, plus roofers and electricians for the home-services group.

And your instinct on the signals is right: for mechanics the evidence mix leans harder on reviews, location, and manufacturer specialization (someone earlier flagged "independent Porsche/BMW mechanic near me" as a perfect example) than on keywords. Same core mechanic across industries, but the specific proof that earns the recommendation shifts — and for trades it's reviews and specialization doing the heavy lifting.

You've nailed the exact conversion moment — the named competitor is what makes an owner's stomach drop, not the score. The score is the hook; "Wappenschmied is being recommended to your customer right now" is what makes them actually do the fixes.

Personal injury lawyers is a great add — highest-intent local search there is, acute need, one client pays for years. Adding it.

On which of the 18 converts best: too early to say with real numbers — this version hasn't been in front of enough people yet. But my honest bet is exactly where you'd guess: highest-ticket categories where the gut-punch lands hardest (lawyers, med spas, HVAC) should convert best, because the cost of being the missing third name is most obviously five figures. That's the thing I most want the data to confirm or kill.

That's a perfect example of the silent-failure problem — and the LLM-generated directory description under a dead company name is almost worse than being absent, because it's confidently wrong and you'd never know to look.

You're right that "here's what an assistant literally says about you right now" is a sharper hook than a score — the reports already lean on that (the named-competitor line does it), but pulling the actual current description the model gives is a stronger, more visceral version. Adding that to the report is worth it. Thanks for the concrete case.

Accountants are already live, actually — planmoon.app/for/accountants — and bookkeeping is a great adjacent add for exactly the reason you said: referral-run, plain-language "bookkeeper near me" searches.

And you're right on the design instinct — the report already leads with the "fix today, takes an hour" list over a diagnosis, precisely because the owner who doesn't act same-day usually won't. That's the whole reason for the speed-to-proof framing.

Same gap, different entry point — meeting small businesses where they already are (WordPress for you, their existing site for me) beats asking them to adopt anything new.

Most surprising underserved ones: the high-ticket trades — HVAC, roofing, med spas. You'd assume categories with that much money moving through them would have someone minding the AI channel, but the sites are often a phone number and a photo gallery, so they're wide open. The boring, expensive industries are the most underserved, not the obscure ones.

Great question, and the answer I keep hearing is: almost completely off the radar. Most owners I've talked to are still in "am I ranking on Google" mode — the idea that a customer might ask ChatGPT "who's a good X near me" and get three names, none of them theirs, just hasn't landed yet. There's no alert when it happens, so nothing forces the realization.

The ones who are aware tend to be either younger owners who use AI themselves and made the leap ("wait, I do this — my customers must too"), or people who've already lost a deal and traced it back. Everyone else is genuinely surprised the channel exists at all, which is exactly the problem — you can't fix a leak you don't know is there.

Curious whether that matches what others are seeing, or if it's further along in some industries than others.

That brand-monitoring "80% mentioned" number is a perfect example of why I don't trust a raw mention rate — for a pre-launch app with almost no web footprint, 80% almost has to be the tool counting loose or hallucinated matches, not real recommendations. Mentioned-vs-recommended is exactly the distinction that number papers over, so good instinct flagging it as implausible.

Property managers / landlords is a great add — high-trust, "who do I actually call" decision, and easy to overlook next to the obvious ones. Adding it.

On starebrain: I'll be straight with you before running it, because I'd rather be useful than just hand you a report. Pre-launch with basically no web presence, the 8-question test will almost certainly come back near-zero — but unlike the ZEAM or Amami results, that's not a "gap" finding, it's just "there's nothing for the model to cite yet." The tool is built to catch businesses that are known-but-not-recommended; yours would be not-yet-known, which is a different (earlier) problem where the fix is footprint, not phrasing. If you want, I'll still run it and post the real result — but the honest read is that it'd tell you more once you've got even a thin trail (a launch page, a few directory listings, a Reddit/PH presence). Happy to do it either way — just say the word.

This is the sharpest methodological point in the thread, and you've named a real limitation. Right now the reports capture sources but not the evidence chain — I list what each engine leaned on overall (e.g. the Amami one cites matomo.org, posthog.com, plausible.io, indiehackers.com), but I don't yet map which source fed which specific question, whether the retrieved claims conflict, or which version the model treated as authoritative. So it's outcome + a source list, not a per-question evidence trail.

Which means you're right that a raw 2/8 can't currently distinguish your three cases — content gap vs. trust gap vs. stale-source. Two of those look identical in my output today: a business could add a flawless FAQ and still lose because the model keeps trusting an old directory or a conflicting profile, and my current test would just show "still 2/8" without telling them why.

Capturing cited-source-per-question and diffing claims for conflict is clearly the right next build — it's what turns "you didn't get recommended" into "you didn't get recommended because this stale profile outranked your own page." The fixed-cohort rerun you describe is the other half; without a stable question set run before and after, I can't attribute a change to the fix vs. model noise. Both are on the list now. If you've actually built evidence-chain tracking against ChatGPT/Gemini, I'd genuinely like to compare notes on how you pin down which source was treated as authoritative — that's the part I expect to be messiest.

Honestly, right now I don't fully — divergence between the two is a real limitation, not something I've solved.

What the reports do today is keep them separate rather than blend them: I run all 8 questions against ChatGPT and Gemini independently and show the split (the Amami report, for instance, was 1/4 on each). So you at least see that a business is invisible on both vs. missing on one — which matters, because losing only on Gemini is a different problem than losing on both.

But that's reporting the divergence, not resolving it. When they name different competitors for the same prompt, I don't yet have a principled way to say which signal to trust or how to weight them — and since neither is deterministic, some of that difference is just noise between runs, not a real engine-level disagreement. Untangling "these engines genuinely retrieve different sources" from "same engine, different roll of the dice" needs running each question several times per engine and looking at what's stable, which I'm not doing systematically yet.

For now the practical answer is: I treat a gap that shows up on both engines as the high-confidence signal worth acting on first, and flag single-engine gaps as lower-priority. Crude, but it keeps the recommendations from chasing one engine's quirk. Tightening that is on the list.

That last question is the one I think about most, because it's the real risk to anything in this space — build something tuned to today's retrieval logic and it's obsolete the next time OpenAI ships an update.

My bet is to stay on the side of the equation that doesn't churn. The retrieval mechanics change constantly; what stays stable is that a model recommends a business when there's clear, specific, citable evidence that it's a strong answer to the question — and it falls back on the few recognizable names when there isn't. Every recommendation I make is about giving the model better real evidence to work with (a services block that actually says who you help, an FAQ answering the real question, verifiable proof), not about exploiting a current ranking quirk. That kind of fix tends to survive retrieval changes because better source material helps under almost any retrieval logic, where anything that games a specific algorithm ages badly.

Where the churn genuinely bites is the measurement layer, not the advice layer — a 2/8 score today isn't directly comparable to a 2/8 six months from now if the underlying model changed. That's why the durable output is the fix list, and the score is really just the hook that gets an owner to read it. If a provider fundamentally rewrote how local retrieval works, I'd have to recalibrate the test — but "publish real proof of what you're good at" is about as future-proof a recommendation as I can make.

This is one of the most useful comments I've gotten — you did real work here, so let me actually engage it rather than just say thanks.

The four verticals are all going in. MSPs got a strong second from someone who ran one for 20 years earlier in this thread, and your HIPAA/SOC2 angle sharpens why they fit — those are exactly the specific, high-MRR buyer questions where a generic MSP site has nothing citable and the model defaults to whoever documented their compliance work. Same logic makes the specialized auto repair one underrated: "independent Porsche mechanic near me" is a real prompt with real intent, and it's a category where a good shop is often invisible because their site is a phone number and a photo. Commercial janitorial and roofing/HVAC both fit the pattern too — high deal size, AI as the first filter, generic copy.

Your two technical points are the part I want to sit with:

The JSON-LD schema fix is a genuine gap in my current reports. I've been recommending human-readable content (FAQ, services block) but not explicitly checking for LocalBusiness microdata — and you're right that areaServed / hasOfferCatalog / knowsAbout are exactly the fields grounding leans on. That's a near-perfect "fast fix, huge parsing leverage" recommendation, since it's often a copy-paste block that makes a site dramatically more machine-legible without a redesign. Adding schema validation to the checklist.

The 3-run citation confidence score is the cleanest answer I've seen to the non-determinism problem I keep bumping into in this thread. A single snapshot has variance I can't currently distinguish from a real signal; sampling each question 3x and scoring 3/3 vs 1/3 gives an actual confidence measure without exploding token costs. That's the version of "rerun the questions" I was hand-waving at — you've given it a concrete, affordable shape. That's going straight onto the build list.

Real thanks for this — the schema check and the confidence score are both things I'll actually ship.

Independent B2B consultants is a smart pick precisely for the reason you gave — their evidence is scattered across site, LinkedIn, directories, talks, and client case studies, so they're a clean natural experiment for owned-copy vs. corroborating-sources. If a fractional RevOps consultant moves on a buyer question only after third-party citations line up, that tells you something the tidier categories can't. Adding it.

On the direct question: yes, freezing a question cohort per industry is the plan, and it's the thing that makes any of the rest coherent. Without a stable set run before and after, a 2/8 → 4/8 move is uninterpretable — could be a real gain, could be the sampling noise you're pointing at. So the cohort gets fixed per industry, then each buyer question gets treated as a mention rate over several runs rather than a single pass/fail, and I change one asset at a time so a shift is attributable. Someone else in the thread framed the sampling as a 3-run confidence score (3/3 vs 1/3); your "change one asset at a time" is the missing other half — without it, even a clean mention rate can't tell you which fix moved it.

The part I hadn't separated cleanly until your comment is brand mention rate vs. citation-source churn as two distinct metrics. Those really are different signals: mention rate says whether you got recommended, source churn says which evidence the model started trusting to do it. You could hold mention rate flat while the underlying cited source quietly shifts from a stale directory to your own page — and that's a leading indicator I'd otherwise miss. Tracking them separately is going on the list. This is the second comment in this thread pushing me toward evidence-chain tracking, so it's clearly the real next build, not a nice-to-have.

Ordering: speed-to-proof — quick fixes that directly answer the exact dropped question go first, bigger pages come after.

Rerun question: not yet, honestly. Right now it's one snapshot — test, report, fixes — with no loop confirming which changes actually moved the result vs. just sounding plausible. That's the next thing to build, though I'd want to run each question a few times per round first since AI answers aren't fully deterministic and I don't want to chase noise.

That parallel actually holds up well — "understood → recommended" and "understood → actually used" are the same gap wearing different clothes. Finding/matching content was the hard problem for a decade; now the model can find almost anything, and the failure mode has moved one layer up, to whether it does something useful with what it found. In your case that's turning info into something a learner will act on; in mine it's turning "the model knows this business exists" into an actual recommendation.

Appreciate you following along — I'll keep posting real reports as they come in, good and bad.

Agree, and it's the same idea someone raised earlier in this thread — the aggregate pattern across reports might end up worth more than any single one. Right now every report is a one-off snapshot for one business, but if I keep the raw data (which questions dropped a business out, which competitor got named instead, across all 18 industries), that starts looking like an actual dataset: which categories of questions are systematically hard for AI to answer well, whether certain industries lean on 2-3 dominant names regardless of who's "better," that kind of thing.

I don't have enough volume yet to say anything real from it — this batch is still small — but it's exactly why I'm doing this one report at a time in public instead of just shipping a black-box score. The individual fixes help one business; the pattern across businesses is the thing I'd actually want to write up once there's enough of it.

Good flag — and you're onto something real. Voice search (Siri/Alexa) and AI search (ChatGPT/Gemini) aren't identical, but they share the thing that matters: people phrase queries as full conversational questions, not keyword fragments. That conversational angle is actually the core of what the tool tests — buyer questions instead of keywords — so we're partly there already.

Where I think your point adds something is the structural side: FAQ content written as literal Q&A, question-form headers, and speakable/FAQ schema that makes it easy for an engine to lift a clean answer. The reports already push FAQs, but the schema angle is worth building in more explicitly.

Curious what specifically you added on your site — was it the schema markup, the question-style content, or both? Would genuinely help me sharpen this.

The three-layer split is a useful way to put it — discovered → understood → recommended. Agree the first two are the easy part; the third is the whole game, and it's exactly where the 2/8 businesses are getting stuck. They're discovered and understood fine, they just don't have citable proof for the specific buying decision.

Your last point is the one I'm most interested in: tracking, across all 18, which questions surface a business vs. which swap in a competitor. That's basically the dataset I'm accumulating one report at a time, and you're right that it may end up more valuable than the pages.

One early signal that pushes against the "AI bias toward local businesses" theory: a metal-art studio — not a local-services business at all — also came back 2/8, same pattern, competitors named in its place. Small n, but it hints the gap is content depth rather than a local vs. non-local bias. Whether that holds as businesses actually add the FAQ/service content is the thing I don't know yet — which is exactly the experiment.

This is a great add, and coming from 20 years running one it carries weight. You've named the exact thing the tool is built on — owners knowing their trade cold but having no idea what buyers actually type into ChatGPT. An MSP owner can be genuinely excellent and still be invisible for "how do I pick an IT provider for a 50-person office," because nothing on the site is written in that language.

MSPs are also a near-perfect fit for the 2/8 pattern: high-trust, high-consideration purchase, tons of small owner-run firms, and site copy that (as you said) often reads like 2005 — so the model has plenty to recognize the firm by name but nothing citable for the actual buying decision.

Adding IT/MSP to the next batch. And if you still know owners in that world, I'll happily run reports on a couple and post the real results — good or bad. That'd be a better test of the hook than anything I could argue for it.

Thanks — you zeroed in on the two things I care most about. The "known by name vs. discoverable for real queries" gap is the whole reason the tool exists; a business can be fully understood by the model and still never get named when it counts.

And yeah, "it takes an hour" is deliberate — a visibility score just tells you you're losing, an hour-long fix tells you what to do about it before you close the tab. If you run one of the 18 industries (or know someone who does), happy to run a real report and post what it finds.

Boutique PR/creative agencies is a great call — and the holding-company thing is a perfect example of the 2/8 pattern: the model defaults to the big recognizable names because that's what it has citable proof for, and the actual boutique the buyer wants gets skipped. Adding it to the next batch.

Your point about context beyond citation is the sharpest thing in here, and it's honestly a layer past what the tool measures today. Getting named is necessary but not sufficient — if the answer files you under the wrong category or shows a stale detail, the citation doesn't convert. That's a real gap in what I'm auditing, and it's the kind of thing your side (actual AI-referred visit data) sees that a static test can't. Worth me thinking about how to fold that in.

And I ran yours, since you dropped the link. Amami came back at 2 of 8 — same pattern as the post's example. ChatGPT and Gemini both describe it accurately on branded searches, but on the unbranded ones ("privacy-focused GA alternative," that kind of thing) they surface Matomo, Plausible, and PostHog — 4 mentions each — with Amami absent. So it's not identity; it's the lack of citable sector proof (case studies, deployment stories, a real comparison page) for the specific migration decision. Ironic given what you do, but the quick wins the report flagged are a privacy/data-access statement, a GA-alternative comparison block, and a dedicated MCP setup page — all fast. Happy to send the full thing if useful.

The ordering is basically a speed-to-proof ratio — what's fastest to ship that gives the model something concrete to cite.

Quick wins go first: things like a services block, an FAQ, or a comparison section, because they're a few hours of writing, not a rebuild, and they directly answer the exact questions the business dropped out on. If a buyer question surfaced a specific gap (say, no pricing or process info anywhere), the fix that closes that exact gap jumps the queue over anything more general.

The bigger moves — case studies, dedicated comparison pages, in-depth guides — come after, not because they matter less, but because they take real time and are more valuable once the cheap fixes are live and the business has proof to draw on. There's also a "what's actually true today" filter: I won't recommend a case-study page before the business can say who a real client was and what the outcome looked like. The fix has to be something they can honestly write, not just a page shape that looks good to a model.

Fair worry, since "18 industry pages" could easily become one template with find-and-replace on the industry name.

The main defense is that the report itself is never templated — it's generated fresh per business from the actual test results: which questions it dropped out of, which real competitors got named instead, what specific gap showed up (pricing, process, sector proof, whatever). That's pulled from live data every time, so even two businesses in the same industry get different reports because their actual gaps are different — see the ZEAM (art studio) and Amami (analytics) reports above, both 2/8 but nothing alike in the specifics.

Where the landing pages themselves could drift is in the industry framing around the report — the intro copy, the example gap described, the language used for "what buyers ask." I keep those industry-specific by actually building each industry's buyer-question set from real scraped language for that trade, not reusing a generic set with nouns swapped. If two industries start pulling from suspiciously similar question sets, that's my signal something's gone lazy and needs a rebuild, not a copy-paste.

The pages are meant to be thin on purpose — the report is where the real specificity has to live, and that's the piece I protect from templating.

Good question, and honestly yes — that's turned out to be a bigger effect than I expected going in.

Most owners start from "am I ranking," which is a Google-era mental model that doesn't map onto how AI search actually works. The reports seem to shift that framing just by showing the mechanic directly: you're recognized by name, you're just not being recommended for the question a buyer actually asks, and here are the specific competitors sitting in your spot. That "known vs. recommended" split does more to change how someone thinks about the problem than any fix on the list — a few people in this thread have said basically that, that the gap itself was the surprising part, not the to-do list.

The competitor-names line does something similar — seeing "Wappenschmied" or "Matomo" named in your place makes it concrete and personal in a way a visibility score never does. It stops being an abstract AI-trend problem and becomes "that specific business is answering a question I should be answering."

I don't have hard data yet on whether that reframing changes behavior (do they actually go make the fixes, or just feel unsettled and move on) — that's the thing I'd want to track next, maybe by following up with people who got a report and seeing what they actually shipped.

Good distinction, and it comes down to volume and intent overlap rather than any single question being "special."

A question becomes its own page when it's asked often enough across an industry's buyer language, and answering it well requires enough depth (process, pricing logic, comparisons, examples) that cramming it into an FAQ entry would bury the useful part. If a question keeps showing up in slightly different phrasings across many businesses in the same trade, that's a signal there's real search volume behind it and it can carry a dedicated page.

Everything else — the narrower questions, the ones specific to one business's niche, the ones that are more clarifying than decision-driving — stays as supporting content like FAQ entries. Those still matter for the AI test (they're often exactly what trips a business up in the 8 buyer questions), they just don't have the volume or depth to earn a standalone page.

Rough rule: if multiple businesses in an industry would all want the same page, it's a page. If the answer is specific to one business's situation, it's a fix I recommend in their report instead.

Both — the scraped data gets it onto the shortlist, but the AI responses decide what actually makes the final 8.

The flow: I pull candidate questions from real buyer language (PAA boxes, Reddit/forum threads, review text), which gives me 20-30 per industry. Then I run them against ChatGPT and Gemini to see how they behave. The ones that make the cut are where the model would have to know the specific business to answer well — not too generic (everybody gets named) and not too niche (nobody does). A question only earns its spot if the AI response shows there's a real gap to surface.

So the scraped data tells me what people ask, but the AI validation tells me which of those questions actually separate a recommended business from an invisible one.

Real language first, then filtered by hand — not just what I'd ask myself, though that's a sanity check at the end.

The raw material comes from actual buyer phrasing: People Also Ask boxes, Reddit/forum threads where people ask "how do I find a good X," and review text where customers explain who they picked and why. That gives me 20-30 candidates per industry. I don't have direct access to private AI chat logs, so search-side query data plus that review language is my closest proxy for how people actually talk when they're deciding.

Then I narrow to 8 by running them against ChatGPT and Gemini and keeping the ones where the model has to actually know the business to answer well. The "what would I ask as the customer" gut-check is really just the last filter to catch anything that reads like marketer language instead of buyer language.

That gap is the whole point, and it's why I think most visibility scores miss it — being known isn't the same as being recommended.

Accountants is a great pick, especially given how trust-driven that decision is. Adding it to the next batch.

Here's a reply draft — answers the question and gives them the real report result like you promised in the post:

Thanks! On the questions: it's mostly real query language — People Also Ask boxes, Reddit/forum threads, review text where people explain who they picked and why — filtered down by hand to 8 that are common but specific enough that the AI has to actually know the business to answer well.

And I ran yours. ZEAM came back at 2 of 8 — same as the example in the post, funnily enough.

The pattern: ChatGPT and Gemini both know you when someone asks by name — they'll describe your metal illustration, album art, and logo work fine. But ask the questions a real buyer asks ("artist for dark metal album artwork," "death metal logo design," "reliable album-art commissions") and you drop out, with names like Wappenschmied, Mario Nevado Art, and Loner Illustration showing up instead.

So it's not an identity problem — it's a "no citable proof for the specific buying decision" problem. The quickest fixes the report flagged:

A short services block on the homepage naming exactly what you do (custom metal album art, dark illustrations, death-metal logo design, merch artwork) — ~45 min, gives AI something direct to quote.
A concise FAQ covering process, deliverables, revisions, timelines, and usage rights — ~1 hour, these were the exact things the album-art reliability question dug into.
Project credits on portfolio pieces linking each artwork to the band and release — verified context is what recommendation answers lean on.

Happy to share the full report if you want it — the bigger moves (a dedicated dark-metal commission page, case studies) are where the real lift is, but those three are the fast wins.

Mostly real query language, with some hand-curation on top. For each industry I'm scraping/reading actual buyer questions from places like PAA snippets, Reddit threads, and niche forums, plus review text where people explain their decision criteria ("looked for someone who specializes in X" type language). That gives me the raw material.

From there I narrow it down to 8 by hand — picking ones that are common enough to matter but specific enough that an AI model would have to actually know the business (not just recognize the industry) to answer well. I'm not fully happy with the rigor of that narrowing step yet, so if you've got ideas for validating question sets at scale, I'm all ears.

It's a blend, but weighted toward real signal rather than my own guesses. The starting point is scraped buyer language — People Also Ask questions, forum/Reddit threads, review text — for each industry. That's the raw pool. I then hand-pick the 8 that show up most consistently and are specific enough that getting recommended actually requires the AI to know something about the individual business, not just the category.

Where it's still me making judgment calls is the final filtering — deciding which 8 are the most representative without it turning into 40 near-duplicate questions. That's the part I'd like to make more systematic, maybe by validating against actual model query logs instead of just search-engine proxies.

Good question, and honestly the honest answer is "both, in stages." I start by pulling real language — Google's "People Also Ask" boxes, Reddit/forum threads where people are actually asking "how do I find a good X," review sites where customers explain why they picked someone, and for industries I know well, actual sales call notes or intake questions. That gives me a rough set of 20-30 candidate questions per industry.

how are you actually checking whether a business is being mentioned or ignored by ai search across those 18 industries?

Thanks, and that distinction is the whole thing for me.

On the 8 questions, you are pointing at the exact trap I try to avoid. If the questions sound like the business's own marketing language, the test is too easy and everyone scores well. So I build them from real buyer language instead, People Also Ask boxes, Reddit and forum threads, and review text where customers explain why they picked someone. That phrasing is usually very different from how a business describes itself, which is the point. A consultant says "strategic advisory," the buyer asks "how do I pick an ML firm."

Your experiment idea is a good one, and honestly close to what I want to test properly. Comparing questions built from business language versus real customer phrasing and watching how much the score moves would show how big that gap really is. Right now I lean on the buyer side by design, but I have not measured the two against each other in a clean way yet.