I generated 100+ articles with AI and got 0 traffic. So I built a tool to fix SEO hallucinations.

Generating a hundred articles without live SERP checks is a common trap. Even with better gap data I would still publish fewer pages with clearer intent match, unique examples, and internal links, then only scale once a handful start ranking. Hallucinated keywords are half the issue. Thin interchangeable pages are the other half.

Very interesting read with plenty of useful takeaways.

I would say quality over quantity.
Who will read these 100 articles and what is the benefit of it?

I’ve noticed the same thing. I used to publish AI-generated content almost as-is, but it didn’t seem to perform well for SEO.

Now I put every article through multiple rounds of review, fact-checking, and rewriting before publishing. It takes more time, but the final content is much more useful and much less generic.

This is the bigger problem with AI-generated content: the model can produce something that looks completely reasonable while the underlying assumptions are wrong.

Grounding the workflow in live SERP data makes a lot more sense than asking an LLM to guess what the “content gap” is. Curious if you’re also tracking whether those gaps actually translate into impressions/rankings after publishing — that feedback loop would be really valuable.

Gaps are the easy half. The half that bit us was measurement: once AI-referred visitors land, most show up as direct or a generic referral, so you can't tell which engine or query actually moved. Without that you're optimizing blind, which is why we split AI referrers out at the source (that's the core of amami.dev).

Thats Great

I ran the version of this where the keyword data was real, and I'd push back a little on where the bottleneck is.

I pulled volume/KD/CPC from an actual keyword tool before writing anything, and only wrote against terms that had verified volume and KD under 30. No hallucinated gaps. Twenty-odd articles later, 28-day numbers: 487 impressions, 10 clicks, average position 49.8.

So the hallucination problem was solved and the traffic problem was not. Position 49.8 is page five. My best-performing term got 27 impressions and 0 clicks — not because the keyword was wrong, but because nine other results sat above me.

What I think actually happened: on a domain with no history, keyword accuracy moves you from "invisible for terms that don't exist" to "invisible for terms that do." Both look identical in analytics, which is why the fix feels like it should work. The binding constraint was age and links, and no amount of better research shortens that.

The one thing that did change: with real data I can at least tell the difference between a bad bet and a slow one. Position 49.8 on a term with real volume is a queue I'm in. Zero impressions on a hallucinated term is a queue that doesn't exist. Knowing which one you're in is worth a lot — I'd just frame the tool as a way to stop wasting writing, not as a way to get traffic sooner.

Curious what your positions looked like on those 100+, not just the clicks. If they were ranking 60-90 rather than not indexed at all, that's a different diagnosis.

The sharpest lesson is that one hundred plausible articles can still be one untested hypothesis. I like that TheNicheGap grounds the rewrite in the live top ten and separates missing content from evidence that anyone wants it, though I would still pair every gap with Search Console impressions and a small control group. That makes the next article a measurable experiment instead of another confident guess.

I made a ByteForward rundown of the full Dev Day drop https://youtu.be/6apxniSQix8

The hallucination point is fair, but the failure mode I'd watch for with a gap tool is a different one: a gap all ten competitors missed may simply be a subtopic nobody searches for. Diffing pages tells you what's missing relative to them, not whether anyone wants it. Do you validate a flagged gap against anything — People Also Ask, GSC "other queries", real volume — before calling it exploitable?

One check on top of SERP gaps: ask the same buyer question in AI Overview / ChatGPT / Perplexity before you write the page. If those surfaces already answer it from other sources, a classic top-10 gap can still be a dead page for the traffic that stops at the summary.

Pew (March 2025): 1–2 word searches triggered an AI summary about 8% of the time; 10+ word searches 53%; who/what/when/why starts 60%. So for AI-era SEO I'd prioritize gaps on full buyer questions, not short head terms — even when the SERP gap looks juicy.

The hallucination point is real, but the subtler trap for a gap tool: a gap between you and the top 10 can also be content nobody ever searches for. For every 'what the top 10 missed', check whether the missed subtopic carries its own search demand — otherwise you're writing the perfect answer to a query that doesn't exist. Large-scale AI content tests point the same way: 53% of articles got zero impressions while 66% of those zero-impression pages were cleanly indexed — the pipeline was fine, the targets were wrong. Gap × demand is the intersection worth ranking; a gap alone is trivia.

Great perspective! Really resonates with what we're seeing in the product space right now.

This is a good example of why pumping out more AI content usually doesn’t fix the underlying SEO problem. Hallucinated facts definitely hurt, but I’d also look closely at search intent, whether the topics actually have demand, and if the site has enough authority to compete for them. I like the research-first approach though, especially if every claim can be traced back to a real source before the article gets published.

This hits close to home. I’m building out SEO pages for a small app right now, and I’m deliberately trying not to fall into the “publish 100 AI articles” trap.
The hardest part for me is deciding whether a keyword is actually worth building a page around before spending time on it. Looking at what already ranks + what those pages miss feels much more useful than asking AI for “content gaps”.
Curious to see how this works on smaller long-tail keywords.

I went through a similar pivot. The key insight for me was that enterprise customers don't buy features — they buy risk reduction. Your pitch needs to be 'here's how we reduce your operational risk' not 'look at our feature list.'

One issue in our Formrelay distribution test was more basic: three public Substack posts loaded normally but had a noindex directive when we checked them. We now separate published, indexable, indexed, and generating visits in our tracking sheet. Your reply that about 80% of your pages were indexed is useful context—it points to a different problem than ours. For the next test, I'd compare a small set of revised pages with similar unchanged pages and track impressions as well as clicks. That should give more useful evidence than another large batch of articles. We don't have a traffic improvement to report yet.

Same lesson applies on the other side of the pipe: the answers ChatGPT and Gemini give about a market aren't stable either — they shift by country, prompt phrasing, and week. "Ask the AI once" is basically an anecdote. Your fix is grounding content against live SERPs; the analogous fix for AI answers is grounding them against repeated live queries and tracking which sources actually get cited. Curious how the gap detection holds up on keywords where the top 10 is mostly Reddit/UGC — that corpus churns fast and "what the top 10 missed" can change weekly.

I also using ai content, but its not completely written by ai, i proofread them and read and check basiic, using keywords in that. i crossed 15k impressions and 150+ clicks in 3 month

Yeah, the “100 AI articles → 0 traffic” problem usually isn’t AI itself. Often the pages target keywords with no realistic ranking path, are too repetitive, or lack strong internal linking and differentiation.

I’d start with GSC: check what’s actually indexed and getting impressions, then group pages into keep, improve, consolidate, or remove. For the ones worth keeping, add real examples, unique insights, and intentional internal links rather than simply adding more words.

The exchange about whether it was the hallucinations or the intent mismatch is the most useful part of this thread, and I would push back on the answer you gave: they are not cause and effect, they are two different kinds of claim, and only one of them was ever testable.

"Intent mismatch" is a hypothesis about a reader you have not met. "The LLM invented the search volume" is a claim about a datum you never checked. When you publish a hundred articles generated from an ungrounded model, both are true at once and neither is separable, because you have no measurement that distinguishes them. The hundred articles are not a sample of a hundred hypotheses, they are one hypothesis repeated a hundred times. That is why the traffic result carries almost no information: a single wrong premise will produce the same silence whether the premise was wrong about intent or wrong about demand.

What grounding in live SERP data actually buys you is not better keywords. It is the ability to be wrong in a way you can see. Once the inputs are observable, a bad article is a failed test rather than an unfalsifiable story, and the second attempt is informative. The ungrounded version cannot fail informatively, which is the real cost, and it is not visible in the traffic number.

Where I come from on this: I founded Piramyd, and the whole reason it exists is that my own first attempt at content had exactly this shape — confident, plentiful, and impossible to attribute.

So do not ask which cause it was. Ask which one you can now disprove. With live SERP data in the loop, if an article still gets nothing, what does that tell you that it could not tell you before?

100 articles to zero traffic is the cautionary tale everyone needs - volume without search intent is just expensive diary writing. we're promoting swapfile.live the opposite way: showing up where people already ask for file conversion instead of publishing into the void. did the hallucinations turn out to be the actual traffic killer, or was it intent mismatch all along?

Disclosure: I run UtilitySEO, an SEO and AEO scanner, so we're adjacent.

Your reply further down is the interesting part: about 80% indexed, but stuck on pages 4 and 5. That usually means Google read the pages and trusted other sites more, which is an authority problem more than a content one.

We live there. UtilitySEO is at DA 3 with 10 referring domains (all spam) and about four search visits a month. Some of our pages cover their topic properly and still sit deep in the results.

That suggests a feature: show the authority of each top-10 result next to the gaps. A gap on a SERP full of DA 80 sites isn't winnable for a new domain. A gap where two of the top 10 are small blogs or forum threads is.

Does TheNicheGap flag SERPs a low-authority site can realistically win?

Hit the same wall with a programmatic SEO test last year — the keywords the LLM suggested looked reasonable, but half of them had no real SERP at all when I checked manually. Grounding the analysis in actual top-10 results instead of model guesses is the right instinct. Curious how you handle keywords where the top 10 is all UGC like Reddit and Quora — does the tool treat those as gaps, or as a signal to skip?

Really interesting case study. The biggest takeaway for me is that AI can generate plausible SEO strategy, but plausibility isn’t the same as search evidence. Grounding content-gap analysis in live SERP data makes much more sense than letting an LLM invent opportunities from its training data. The distinction between “missing content” and “unfulfilled search intent” is especially important.

The live SERP approach makes a lot more sense to me than asking an LLM to guess what Google is rewarding. I'd like to see how consistently the recommendations hold up across competitive keywords rather than just obvious gaps.

The core insight here is that you're solving a signal verification problem, not just a gap-detection one. AI hallucinating gaps is the visible failure mode; the harder one is when a "gap" is actually a topic the top 10 results deliberately don't cover because it's not what searchers need. If you reverse-engineer intent from what's ranking, you're crowdsourcing the verification that matters - the market already told you what the answer shape is. A measurement system for content gaps works when it grounds itself in observable outcomes (what Google ranks, what searchers click) rather than plausible text patterns.

Similar experience here, with a different diagnosis. We published 222 AI-written comparison articles; 93% got zero views and the site ended up mostly deindexed. When we dug in, the biggest problem wasn't keyword choice: our sitemap had 750 URLs and 560 of them were thin tag and category pages, plus near-duplicate articles. Noindexing the thin pages (750 to 124 URLs) and rewriting a few pages with verifiable data mattered more than producing more content. Curious whether your tool flags index bloat, or only SERP gaps?

Zero traffic on 100 pages can also mean Google never took them in, which no SERP data will show you. I had the same flat line and was sure it was my keywords. Then I compared my sitemap against Search Console: of 281 URLs, 256 were ones Google had never seen. It was an indexing problem, not a ranking one, and I had spent weeks tuning pages nobody could find. How many of your 100 show as indexed right now?

This tracks with what I've seen building my own AI video tool. The model sounds confident whether or not it's right. We had to ground every generated claim in what the product actually does, otherwise it's just plausible text, same issue as your SEO example. Do you validate against the live SERP data before showing gaps, or trust the scrape and let the LLM summarize on top?

I feel like I have followed the same trap with having an LLM suggest approaches to develop AI content. This looks like a good tool to try and bridge the gap with real data.

Have you been creating content based on the gap analysis your tool has done for you personally? What kind of success have you had vs the LLM approach?

Interesting timing. I’m testing almost the opposite approach right now — publishing very few pages and letting Search Console show me what Google is already testing the site for.
Today I saw “resume skills” queries getting impressions, so I built a dedicated page around that instead of guessing the next keyword. Curious to see how this compares after a few weeks.

Have early tests shown that TheNicheGap's live SERP gaps actually change what users choose to publish, or are people mainly validating that the recommendations look more credible than LLM guesses?

The biggest issue with AI SEO for me is exactly what you mentioned: AI can generate a very convincing strategy without actually validating whether the underlying search demand exists.
I like the idea of grounding the analysis in live SERP data rather than letting the model guess what the gaps might be. I'd be especially interested in how you distinguish a genuine content gap from a topic that the top results simply don't mention because it isn't important to search intent.
The UI looks interesting too. Curious to see how accurate the gap recommendations are on difficult, competitive keywords.