I handed my content marketing to agents. 6 months later organic traffic is 34x.

The 50–100/mo band finding is useful, but for agency work I’d add one client-facing guardrail: separate discovery targets from conversion pages. A low-volume query is a great wedge when it maps to a problem the client actually sells; otherwise you’re just accumulating cheap impressions. The cohort reporting plus explicit retarget/retire rules feels like the part worth turning into a repeatable monthly retainer deliverable.

The band data is the part I'll be stealing. Everyone says "go long tail" and nobody publishes median position by volume band.

One thing I'd love to know from a site your size, because I can't test it at mine: are the assistants citing you? You've got 23,000 pages indexed on queries people actually type. If you took ten of your best-performing articles, asked ChatGPT and Perplexity the question each one targets, and checked whether your domain gets named — that's a different scoreboard from Search Console, and at your scale you'd be one of the very few people who could actually answer it.

I ask because what I keep finding on small sites is that the assistant answers with a directory or a competitor even when the page ranks fine in Google. Ranking and being cited come apart more than I expected. At 23,000 pages the sheer coverage might solve it for free — or it might not, and that would be the more interesting result.

Wait, so the big change was just going after smaller keywords? That's interesting. Would that work on a tiny site too?

Genuinely impressive results. One thing I'm trying to square though: to land at median position 14 on 1,000-5,000/mo terms, the domain must have had real authority already, links from pages with traffic, or existing click history in the topic. A brand-new site publishing the same articles would mostly sit in "discovered, not indexed," not page two.

So what was the domain's starting position underneath the flat traffic line? Existing backlink profile from the 22 years, or was most of that pre-Feb traffic branded/navigational to one page? Trying to understand how much of the 34x is the agent pipeline versus pre-existing authority the content finally had something to deploy through.

The point about checking search intent before creating a page is really important. For smaller tools, targeting a very specific query can sometimes be more practical than competing for broad keywords. For example, someone searching for a way to combine names already has a clear intent, so a focused this resource can be a relevant solution. Do you think adding a search-intent check before publishing would reduce the number of pages that end up with zero impressions?

Nice

The deletion side is the part almost nobody writes about, and it's the bit that made this click for me. Ratio of live-but-dead pages is a signal a spam system can actually compute, and if you only ever add, that ratio only ever goes up.

The volume band point matches what I've seen. 50-200/mo is the sweet spot on new-ish domains. Once you try to muscle into 1k+ terms without the internal linking and topical depth to back it, position 11-14 is exactly where you end up, and position 11-14 pays nothing. Better to own 30 boring queries at #3 than fight for one big one at #12.

One thing I'd be curious about: on the retarget queue, are you letting the agents propose the pivot keyword and just approving/rejecting? That's the shape that worked for me when a similar queue got out of hand — I stopped trying to pick the redirect target and started grading a shortlist of three the agent surfaced. Cut decision time by maybe 10x without actually handing over the call.

The keyword-band data is the most useful thing in this post, and it matches what I see across the companies I back: cheap rankings on small terms beat expensive rankings on big ones almost every time. What I would want next is revenue per band, not traffic per band. A 50 to 100 search term you own is still worth nothing if the searcher has no intent to buy, and I have watched founders turn 30x traffic into flat MRR more than once.

The first two months were not a quality failure. 66% of the dead pages were indexed; they just had no demand and no shot at the top 10. A gate before generation — volume band plus whether you can actually beat that SERP — is cheaper than shipping into a vacuum and hoping the 28-day redirect cleans it up.

The deletion loop is the part worth copying. 200 a day without a standing noindex on zero-impression pages is how the dead-page ratio climbs, and that ratio is what a spam system can see. Volume is cheap. The human sign-off on what gets cut is the control. Handing that to the same agents that generate is how you trade the trust those link requests were reacting to for throughput.

That zero-impression rate is the useful warning: indexed is not the same as discoverable. The tiered targeting plus a deletion loop feels like the right operating system—I’d also sample whether AI answers surface the pages that win long-tail search, since clear one-question coverage tends to travel well there. Really strong example of treating content as a portfolio, not a pile.

The useful lesson here looks less like agents writing 200 articles and more like a feedback loop for choosing the next unit of work. The 53% to 25% zero-impression change gives a clean quality signal. I would also track impressions per article by target band and deletion or retarget decisions, so traffic growth is not hiding a growing dead-page backlog.

Solid point. Validating user demand before getting too deep into the architecture saves so much time.

Hey, thanks for this article. Can you add more details on the removal? How does your script know the number of impressions?

The before/after makes a strong case for target selection, but it also points to a useful cohort metric: distribution of pages by first-impression date and survival after 28/56 days. If deletion or redirect is automatic, I’d keep a tombstone log and periodically sample removed URLs for backlinks or nearby demand, so pruning doesn’t erase assets that are merely slow for non-query reasons.

The drop in zero-impression pages from 53% to 25% after moving from 1,000–5,000 searches toward the 50–100 band is a strong controlled lesson. I would separate the daily report into targeting, index status, and page-type mismatch so the 2,418 retarget items do not all look like the same problem. Do you expect the 56-day noindex rule to vary by topic authority, or is one expiry window easier to keep honest?

"Running agents or they're running you" is uncomfortably close to my week.

I had a team of agents (forked from Paperclip) creating content, posting, and setting up outreach. Every report said it was going well, and all that activity produced 14 unique views.

The problem was exactly what you describe: they were grading their own homework. ended up scraping that team of agents and my new one is much simpler. It's basically scheduled tasks in Linear plus one manager agent whose only job is to verify whether or not the work was useful (as well as a Telegram agent I can use to communicate with them from my phone).

The 53% to 25% zero-impression drop makes the targeting lesson much more actionable than the headline traffic number. I’d keep a small holdout cohort by volume band and publish month, so you can tell whether a change improved the pipeline or just coincided with seasonality and domain aging.

The zero-impression rate by publication month is the number I'm taking from this. I run a SEO and web design agency, MBStudioz (https://mbstudioz.in), working mostly with local businesses in India, where the sites are small and have little authority. Your result that the 50 to 100 searches band beat the 1,000 to 5,000 band on position and top-10 rate is a useful reminder to check demand before choosing what to write.
A question from the small-site end: your cull works because thousands of other pages carry the site. For a client with 30 pages, where some pages exist to convert rather than rank (pricing, contact, a service page), would you still apply a zero-impression expiry, or exempt pages by purpose?

The open source route is interesting but tricky for B2B. Your community contributors want features that help developers, but your paying customers want features that help business operations. Those are often different things.

This one hit close to home — the rank-by-volume-tier numbers (8.6 median at 50-100/mo vs. 14.4 at 1k-5k) are exactly the kind of thing that should be driving keyword selection, and it made me go audit my own pipeline.

I run BlogPilot, a WP plugin that's supposed to score keywords by difficulty + volume before queueing them for content. Turns out mine had been silently broken for months — it was calling a DataForSEO endpoint that doesn't actually exist, so every keyword defaulted to a fake KD of 30 no matter what. I'd effectively been picking keywords blind, same trap this post describes just from a different failure mode.

Since reading this I've: fixed the endpoint so scores are real, added volume-tier badges so a near-zero-search-volume keyword gets flagged instead of silently queued, and shipped a bulk re-score so the whole backlog gets corrected instead of just new keywords going forward. (For anyone curious what it does: https://superappsdomain.com/apps/wp-blogpilot-wordpress-plugin/)

Appreciate you putting real numbers on the volume/rank tradeoff — it's the push I needed to stop trusting a black-box score I'd never actually verified.

The homepage concentration is the part I'd watch most. I'm Benny, founder of AI Undetectable, an AI rewriter and humanizer: https://aiundetectable.com

Our September 2023 GA4 export shows 38,305 users, but 87% of sessions landed on the homepage. Organic-search sessions fell from 47,636 that month to 2,080 in June 2024. In the three months ending September 15, 2026, the blog received just five Google clicks.

That's why I'm not treating more articles as the recovery plan. I've started by fixing legacy URLs and consolidating overlapping guides. Too early to claim a ranking improvement.

On your 56-day rule: do you make exceptions for pages with useful backlinks or assisted conversions, or is zero search impressions enough to retire them?

Written with AI assistance; figures are from my Analytics and Search Console exports.

I dont know that there's that much magic in this - if you keep posting pages until you start getting clicks from Low KDs thne you're scaling authority. Or you're eventually aligning topics in pages that aligned with the domains old authority if it has any (?)

Its still scaled content though - I wonder if reading from the other comments if thats the aim - to "make scaled content" an acceptable practise or a way to generate a PBN or an affiliate site for free?

The age of the domain and the roulette still play and reward the wins mihgt be the 'magic" part....

The deletion logic is the part nobody talks about and you're right that it matters more than publishing. The ratio of dead to live content is exactly what spam systems read. What's your current threshold for deciding something is worth retargeting vs just cutting permanently?

Wait for next core update bro, you will get the good results

The low-volume targeting result (median position 8.6 at 50-100 searches vs 14.4 at 1,000-5,000) is the most useful number here for a brand-new domain. I'm at the opposite end: a small number of hand-checked articles on a domain with almost no authority, so this changes how I'll pick queries.

One question for a topic like mine (regulatory): at 200 articles a day, how do you keep factual claims accurate? I recently dropped a planned "fines by country" article because several sources quoted three different maximum fines for the same country, and I didn't want to publish a number I couldn't trace to the law itself. Does your pipeline check claims against primary sources, or is the long tail mostly low-risk how-to content where a mistake costs little?

The "found out from the log" line on the 1,601-article cull is the part that'd worry me more than the 53% dead rate ever did. Not because the deletion logic is wrong — it sounds right — but because a script that prunes on a schedule and a script that's silently stopped pruning look identical from the outside until you go check. Do you have anything that tells you the job ran and did nothing, versus just... didn't run? I've been stuck on a version of this same problem (a reconciliation job that's supposed to promote stuck records) and the failure mode that got me was a dead job and an empty queue producing the exact same dashboard.

The April data is the most valuable part of this post — 53% zero-impression with 66% of those pages actually indexed kills the 'Google punishes AI content' story and replaces it with a simpler one: when production is free, the only constraint left is demand, and demand lives in the smaller keyword bands. The 50-100/month band beating the 1k-5k band on position, CTR and top-10 rate suggests you haven't found the floor yet — have you tested the 10-50 range, or is there a volume below which a good article's fixed cost stops clearing? That's the number everyone scaling AI content actually needs.

The underrated finding here is that the agents were never the failure point. Same pipeline, same agents, and the only change that moved the numbers was what you pointed them at.

That matches what we see operating agent systems for a living: the generation layer commoditizes fast, and the judgment layer - what to aim at, when to stop, what counts as good - is where the leverage stays. "Cover the top 10 and land at position 11" is exactly the kind of spec error a pipeline will happily execute a thousand times.

Also respect for publishing the 53% dead rate. Most people only post the hockey stick.

34x organic traffic solo? That's insane. The fact that you didn't write a single article yourself and still pulled this off shows how much the game has changed. Most people are still debating whether AI content works — you've got proof. What's your quality control step look like at 200 articles/day?

Really appreciate the transparency here. Real breakdowns and honest reflections are super valuable for the community.

This is the part most teams miss: "I've never done outreach and I've never run a link building campaign. Adding a blog was the whole intervention." But the blog isn't the intervention - the deletion discipline is. Most setups publish forever and accumulate a dead weight of pages that all share the wrong signals with Google. You're deleting 53% in April, then 35%, then 25% - that's the tuning loop that made the 34x work. Without deletion as a measurement-driven decision, volume just becomes noise that confuses the algorithm about what your site actually is.

the deletion pipeline might be the most underrated part of this whole setup. everyone obsesses over publishing volume but the dead-to-live ratio is what actually matters for site health. curious whether the 28-day threshold was something you arrived at empirically or more of a gut call

Building on the retarget-queue point above - I think the fix is to stop treating it as a per-article decision and score it before it ever reaches you, using the variable you already proved in your own data. Only 11% of dead pages had an adjacent 100-2,000 keyword, and your best band is 50-100 searches at 68% top-10, so the agents can attach to every queued article the best adjacent keyword under roughly a 500-search ceiling plus the median position your existing pages hold in that band. Anything with no adjacent keyword under the ceiling never reaches your desk at all; it goes straight to the 56-day noindex path, which is where it was heading anyway. Then cluster what's left by topic so you're issuing twenty cluster verdicts a week instead of scrolling 2,418 URLs - you keep the judgement call, you just stop paying the per-item cost of it. The deletion loop is the genuinely underrated half of this write-up, by the way: the ratio of dead to live pages is the thing a spam classifier can actually measure, and most people copying agent pipelines only ever add. Two things I'd like to know: are the zero-impression articles concentrated in particular topics or spread evenly across the 23,000? And since ad revenue tracked clicks almost 1:1 through 34x, have you seen RPM soften as you push further down the tail, or does a 50-searches-a-month query monetise about as well as a 2,000 one?

The deletion piece is the one most people skip, and your numbers show why it matters. If you only ever add, dead pages dilute the signal. The 56-day noindex pipeline running automatically is the bit I'd steal first.

The thing about the retarget queue — 2,418 articles waiting for a decision — that's the constraint I'd focus on. Right now it's the only part of the operation that still scales with your time, which makes it the only real bottleneck left.

What's the typical search volume range for topics in the retarget queue? If most are just slightly off on keyword tier, a templated reclassification rule might clear it without adding human judgment to each one.

The volume-band table is the most useful thing I've read on this in a while. One data point from a much smaller site: our template pages sat at positions 26–52 for searches like "italian restaurant website design", and moving down in volume wouldn't have helped, because the top 10 were galleries and inspiration posts, not tools. People wanted ideas, not a builder. (Disclosure: I'm building Mythex, an AI app builder; that's our Search Console.)

So a question on the pipeline: before writing, does it check what kind of page ranks (articles vs product pages vs galleries)? Your 11% adjacent-keyword finding makes me wonder whether some of the 1,412 dead pages were a page-type mismatch rather than a size problem.

Your own 1.9 pages per person says more about the retarget queue than any keyword metric. An audience that never browses can't be walked around with internal links or another article, so the retarget that has a chance is aimed at the URL that already gets the click: take a query where you're sitting 11-20 and answer it on the page that's already ranking instead of spawning a new one. That also shrinks the 2,418, because the pages worth touching are the ones Search Console already shows one position away from a query with demand. I see the same shape in long-tail sessions: people arrive with exactly one question and leave the moment it's answered.

The scale is impressive, but the retarget queue sounds like a routing problem more than a writing problem.

I’d group the 2,418 pages using signals you already collect—topic cluster, keyword tier, article age, index status, first-query impressions, and SERP volatility. Review a small sample from each group, then give that group one route: target a reachable long-tail variant, consolidate into a stronger existing page, or retire it.

That keeps the final call human, while turning thousands of page-by-page decisions into a handful of repeatable policies. I’d also track whether each retargeted group earns impressions or clicks after 28–56 days, otherwise the queue will refill without teaching the system anything.

The deletion schedule is the part I want to steal. I am at the opposite end of the authority spectrum and it applies even harder there.

DA 3, ten referring domains all scrapers, four visits a month. Every page at zero impressions is actively diluting the crawl budget that might get one good page seen. You can afford a 1,601-page cull when thousands of others pull clicks. I would be cutting from maybe 40 pages, so the discipline has to start before publication.

Your April-to-June zero-impression drop, 53 to 25 percent, by moving the keyword tier down is the cleanest experiment I have seen on targeting versus quality. Same pipeline, same agents, same domain. Only variable is which queries you aimed at. That falsifies the objection that AI content is penalised, because the content did not change.

Does your 28-day retarget script check whether the page was crawled, or just whether it got impressions? Never fetched is a different problem from fetched-indexed-and-ignored.

The retarget queue is the interesting question, and there is a second axis you can sort it on that costs nothing to compute.

First split the 2,418 by impressions, because two different defects are sitting in one bucket. Zero impressions is a targeting miss and your existing tiering already fixes it. Impressions with zero clicks is a title and intent problem, and retargeting the keyword will not touch it.

Then sort the rest by whether the answer can be inlined. We measured this on two sites this month. One carried 87,400 Copilot citations over 90 days and zero sessions attributable to an AI source, because every top cited page was a review or a pricing comparison, which an answer engine reads out in full and the reader is finished. The pages that turned a citation into a visit were the ones whose value cannot be carried inside the answer: setup guides, live comparisons, anything with a tool or an input on the page.

So for 2,418 articles, rank by how inlineable the answer is and work the non-inlineable ones first. An inlineable page sitting at zero will still be at zero after you retarget it, because the click was never available to win. That also reads on your own numbers: 12.4 pages per person down to 1.9 is what one-answer-per-page looks like, and it is fine while ads are the monetisation, but it means every page has to earn its click on the way in rather than from the page next to it.

Caveat: two sites, neither publishing at your volume, so it is a lead and not a law.

The other thing your post surfaces and does not measure is the inbound. Link exchange requests from listed companies arriving daily, after years of none, is a channel you have opened and are not instrumenting. If you want a free read on the channels the blog is now earning attention on, contentmation.com/score runs a pass with no login. At your scale it will tell you nothing new about search, which is most of your traffic, so it is only worth the minute for the rest.

I'm testing And copy pasted my agent to see if it works. for my b2b

Two things stood out. First, "the tier you aim at decides the rank you get" is a cleaner way of saying what most SEOs learn the hard way: scaling output without tiering is just burning index budget. Your April to June pivot, from 53% zero-impression pages to actually ranking, is a great case study in that.

Second, the link exchange requests from listed companies feel like the real moat forming. The 34x traffic is the headline, but inbound interest from much bigger companies means the library is becoming a reference asset that compounds on its own. I'd track that inbound-request rate as its own KPI - it's a leading indicator that the content has crossed from "indexed" to "authoritative."

This is the measurement visibility problem at scale - you handed off the work but now have to measure what the system is actually doing. The 34x traffic number is powerful because it's observable, but most teams using agents for content never set up the instrumentation to catch what worked. Measuring agent outputs backwards from actual user behavior is the only way to know if you're running agents or they're running you.

The deletion system may be the most valuable part of this, but I’d keep a small control group before trusting the 28/56-day rules completely.

Leave perhaps 5–10% of zero-impression pages untouched for another 60 or 90 days, grouped by keyword tier and index status. Then you can see how many would have surfaced naturally versus how many remained dead.

Without that holdout, pruning and growth happen together, so it is difficult to separate the value of deletion from the improved keyword targeting. It would also tell you whether different tiers deserve different expiry windows.

The 2,418-page retarget queue is more interesting than the 34x traffic. How do you decide which pages deserve another target versus deletion?