Read left to right. We were sitting comfortably around 280–330 impressions a day in mid-August. From 18 August the line starts walking downhill — not a cliff, a slope. By 22 August we were at 190. By 29 August we bottomed out near 103 and stayed flat there for three days.

That is a 69% loss of AI visibility, and the shape of it is exactly why we missed it. A cliff triggers alarms. A slope looks like seasonality. Late August, people are on holiday, traffic is soft everywhere — we told ourselves that story for about six days before someone asked why organic clicks had not softened by anything like the same proportion.

That was the tell. Classic blue-link impressions were down about 8% over the same window, which genuinely was seasonal. AI impressions were down 69%. Two numbers that normally move together had decoupled, which meant something specific to the AI surface had broken.

What actually broke

On 18 August we shipped a redesigned article template. It looked better. It also did three things we did not think about:

It moved the answer below a lazy-loaded component. The new design opened with a "Key takeaways" card that rendered client-side. In the server HTML, the first thing after the H1 was an empty div. The paragraph that actually answered the page's question had been pushed down and, for anything reading raw HTML, effectively buried. Passage-level retrieval scores chunks, not pages, and a disproportionate share of citations come from the opening stretch of a document. We had emptied ours.

It turned our FAQs into client-fetched accordions. Previously the FAQ answers sat in the HTML and were hidden with CSS. The new component fetched them on click. A human clicking sees the same thing. A crawler sees an accordion full of nothing. Our FAQ blocks had been quietly earning a meaningful share of our citations, and they went dark.

It added a redirect hop. The redesign came with a path change from /blog/ to /resources/. Our older URLs already had one legacy redirect on them, so the new rule stacked a second hop on top, and a handful of syndicated inbound links made it three. Googlebot tolerates long chains without complaint. AI retrieval crawlers are far stricter — they abandon at around three hops and, critically, they fail silently. No error, no report, no line in Search Console. The page simply stops being eligible.

None of these were content problems. Every page was as good on 29 August as it had been on 17 August. We had just made all of them unreadable to machines, in three separate ways, in a single release.

The gradual slope makes sense once you know the cause: the decline tracked Google's recrawl of our library. Pages dropped out as they were re-fetched, one batch at a time, over about thirteen days. That is what a template regression looks like on a chart, and it is worth committing to memory — a slow, even decay across an entire site usually means something structural, not something editorial.

What we changed, and how fast it came back

We shipped the fix on 1 September. Three changes, all boring:

• Server-rendered the opening answer paragraph so it sits in the raw HTML immediately after the H1, in a self-contained sentence that reads correctly if you quote it alone.
• Moved FAQ answers back into the HTML, hidden with CSS rather than fetched on interaction.
• Collapsed every redirect chain to a single hop and repointed the legacy rules directly at final URLs.

First movement showed up on 3 September: a spike to around 358 impressions, higher than the pre-drop peak. That is not magic, it is recrawl — a batch of pages came back eligible at once. It settled back to 205 the next day, then climbed and stabilised in a 250–350 band for the rest of the month, with the usual weekday sawtooth.

Three days from fix to first signal. Roughly two weeks to a stable band. That timeline is specific to recovering something you already had: the authority and the third-party consensus were untouched, so once the mechanical blockers cleared, our old position came straight back. Building that position from zero is a four-to-eight-week job on ChatGPT and Google, faster on Perplexity, which weights recency heavily and refreshes its index aggressively.

Why this happens to good sites constantly

The uncomfortable finding underneath all of this is that AI citation and Google ranking have come apart. Ahrefs measured that in July 2025, 76% of pages cited in AI Overviews also ranked in Google's top 10 for the same query. By February 2026 that figure had fallen to 38%. BrightEdge's separate analysis put the overlap lower still, around 17%.

Two things follow. Your rankings no longer protect you — we held our positions throughout the drop and it made no difference. And your competitors' rankings no longer block you, which is the good news for anyone starting from a weak domain.

It also means each engine has to be treated separately, because they do not share an index. ChatGPT leans heavily on third-party directories and review sites; Profound's citation analysis found directories accounting for roughly 48.7% of its citations. Perplexity leans on community and recency — the same analysis found Reddit making up about 46.7% of its most-cited sources. Google AI Overviews and AI Mode run on Google's own index. You can be highly visible in one and completely absent from another, and only find out if you check each one.

The checklist

This is what we run now, quarterly on the whole site and before every template release. It is ordered deliberately: Parts 1 and 2 are where sudden losses come from and they are cheap to fix, Parts 3 to 6 are where durable position comes from and they are not.

Part 1 — Access: can the machines reach you?

• robots.txt explicitly allows retrieval bots: OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User.
• You have made a deliberate, documented decision on training crawlers (GPTBot, ClaudeBot, CCBot, Google-Extended) — separately from the retrieval ones. Blocking training is a legitimate choice; blocking retrieval removes you from AI answers entirely.
• Deprecated agents (anthropic-ai, Claude-Web) removed so the file is not carrying dead instructions.
• Your CDN, WAF or bot-management layer is not challenging AI retrieval bots. Check this separately — it runs before robots.txt and overrides it. Aggressive default bot-fight settings are the single most common cause of invisible AI blocking.
• Server logs confirm the bots arrive and get 200s, verified by reverse DNS rather than trusting the user-agent string.
• Redirect chains are one hop maximum.
• Primary answer content is in the server-rendered HTML, not injected client-side.
• Tab and accordion content is present in the raw HTML, not fetched on click.
• sitemap.xml current, referenced in robots.txt, returning 200.

Part 2 — Extraction: can one clean passage be lifted?

• Every page answers its core question in the first 40 to 60 words, in a complete sentence that survives being quoted alone.
• That opening answer does not depend on the headline, an image or a previous paragraph to make sense.
• Each H2 is phrased as a question a buyer would actually type or say.
• The first sentence under each H2 directly answers that H2.
• Paragraphs run 40 to 80 words.
• Numbers, dates and prices appear as HTML text, not only inside images.
• Comparisons use real tables with proper headers; lists use real list markup.
• Marketing hedges removed. "Arguably one of the leading solutions" cannot be cited. "Cuts onboarding from 14 days to 3" can.
• One page per distinct question, rather than five answers buried in a mega-guide.

Part 3 — Entity clarity: does the machine know who you are?

• Organization schema with legal name, logo, founding date and sameAs links.
• Article schema with a named author; Person schema with real credentials.
• FAQPage schema only where visible FAQ content actually exists.
• Brand name, one-line description and category stated identically across your site, LinkedIn, Crunchbase, G2, Capterra and any Wikipedia or Wikidata entry.
• Pricing and feature names match everywhere. Conflicting facts across sources cause models to drop the claim rather than pick a side.
• Clean, query-matched slugs.

Part 4 — Consensus: what do other sources say about you?

• You have run your priority prompts and logged which domains get cited. That list is your real target list — not a generic outreach list.
• G2 and Capterra profiles complete, claimed and recently reviewed.
• Presence in the roundups and "best X" listicles that already appear in your category's AI answers.
• Genuine participation in relevant community threads. Not astroturfing — models and moderators both catch it.
• Original data or benchmarks that give other people a reason to cite you.
• Named expert commentary in the five to ten publications that dominate your category's answers.
• A documented process for correcting factual errors AI engines repeat about you: publish the correction on a source they already cite, then re-check the prompt in two to four weeks.

Part 5 — Freshness: are you still true today?

• Visible "last updated" date matching dateModified in schema.
• Priority pages reviewed on a fixed calendar.
• Updates are substantive. Bumping a date without changing content is detectable and does not work.
• Any statistic older than 18 months re-verified or replaced.

Part 6 — Measurement: would you notice a drop in 48 hours?

• A fixed library of 30 to 50 buyer-intent prompts, written down and version-controlled.
• Those prompts run on a cadence across ChatGPT, Perplexity, Gemini, Claude, Copilot and Google AI Mode — each has its own index, so each needs its own measurement.
• Search Console → Performance → Generative AI features checked weekly, and compared against your blue-link trend. The gap between them is the early warning.
• Referral traffic segmented by AI source in analytics.
• Server-log monitoring of AI bot hits, with an alert if crawl volume drops sharply. This is the check that would have caught ours on day two.
• Mentions and citations tracked separately — being recommended without a link is a different problem from being linked without a recommendation.