A page that stopped showing up in AI answers did not simply expire. Once the drop holds up across runs, find the stage of the answer-building process where it fails, because the stage decides whether you repair the page or replace it. Four stages can fail: the page is never fetched, it is fetched but never selected as evidence, it is selected but the answer states something wrong about you, or the question buyers now ask is not the question the page answers, and only that last one calls for a rewrite. The first three are repairs, and repairs are cheaper and easier to measure.
The stakes are the shortlist. 51% of software buyers now start research with an AI chatbot more often than with Google, and G2's June 2026 survey found AI chatbots and review sites are now the two biggest influences on B2B shortlists. Buyers also arrive knowing most of their options, since 6sense found the average buying-group member knows 75% of the vendors on the shortlist by the time it is filled. Guessing wrong is expensive in both directions, and the top-ranking page now sees a 58% lower average clickthrough rate when an AI answer sits above it.
Why did the page stop showing up in AI answers?
Because one of four stages failed, and each fails independently. An answer is assembled by fetching pages, selecting passages from what was fetched, composing claims out of those passages, and matching the result to the question asked. Our walkthrough of how AI reads your content covers the machinery.
Stage 1: the page is never fetched
The symptom is absence from your own server logs: no AI user agent requests the URL, or one gets a shell whose content only appears after JavaScript runs. The crawler study Vercel ran with MERJ found that OpenAI's and Anthropic's crawlers fetch raw HTML and do not render JavaScript, while Google's Gemini leverages Googlebot's infrastructure, which does. Google is explicit that pages must be indexed and snippet-eligible to appear in its AI features, and OpenAI publishes four relevant user agents you can check your logs and robots rules against.
Stage 2: fetched, but never selected
The symptom is presence without participation: your URL is fetchable and indexed, and competitors' passages get quoted instead. Seer found 87% of 500 SearchGPT citations matched Bing's top results, while Ahrefs found 38% of AI Overview citations rank in the organic top 10 across 4 million AI Overview URLs in March 2026. Selection rewards extractable evidence: the GEO study published at KDD 2024 found statistics, quotations and citations lifted visibility in its benchmark.
Stage 3: selected, but the answer gets you wrong
Your page can be cited in an answer that still describes you inaccurately, because the claim is assembled from several sources at once. Profound found more than 4 in 10 AI citations point to sources brands do not control, so a stale roundup or an old forum thread can outweigh your own copy. Ambiguous identity compounds it, so use one name and one description everywhere; that helps engines attribute each claim to the right company.
Stage 4: the question moved
The page is fetched, selected and accurate, and still loses, because buyers stopped asking what it answers. A page built around a workflow nobody runs now cannot be repaired into relevance. This is the only stage where a rewrite is the correct instrument.
How do you tell which stage failed?
Re-run a fixed set of buyer questions and watch where the page drops out. Keep the set stable across quarters, because a moving question set cannot tell you whether the page changed or the market did. We use a few weeks of answers per question as a working default, in two or three phrasings and on two engines, recording who was cited and which URLs carried each answer. The right number depends on the variance in your own sample.
Single runs prove nothing. A 2026 variance-components study found run-to-run noise large enough to swamp real differences in small samples, and SparkToro's research reports AI engines highly inconsistent when recommending brands.
Then read the record for a fingerprint.
Your URL never appears, and a JavaScript-disabled fetch returns an empty shell: stage one.
Your URL is fetchable, and competitors' passages get quoted instead: stage two.
Your page is cited, and the surrounding sentence is wrong: stage three.
The answers are reasonable and address a different problem: stage four.
Which repair matches which stage?
Match the intervention to the stage, and stop there. Stage one is mechanical: server-render the content an engine needs, and check your robots rules and indexability against the documented user agents. Cloudflare recorded GPTBot's share of crawler requests rising from 2.2% to 7.7%, and that evidence sits in your own logs.
Stage two is an evidence edit on the page you already have. Put the number, the named source and the direct answer into the section that already covers the question, and keep a heading structure an engine can segment. That edit makes an existing answer easier to extract, and the C-SEO Bench result below explains why extraction alone rarely moves selection. Google's structured data documentation describes what markup does and does not do, and its people-first content guidance still describes the substance being selected for.
Stage three is a claims repair that mostly happens off your own domain. Correct the fact on your page, then correct it where engines are reading it: the directory listing and the comparison article carrying the outdated description. Keep substantiation for every claim, which is the standard the FTC's advertising guidance already sets. Watch the persuasion dial: scarcity and exclusivity framing measurably reduces how often a model recommends a product.
When is a rewrite actually the right call?
A rewrite is right when the question itself changed. The signal is in the answer text: engines are addressing a problem statement your page does not contain and citing pages that frame the topic differently. That is a different article wearing the same URL.
Two secondary cases qualify. A page built for a product you no longer sell has nothing left to repair. A page whose central argument you now believe is wrong needs replacing, because a corrected sentence inside a flawed frame produces an incoherent source.
A rewrite also resets whatever history the page had. Trading a known position for an unknown one is defensible when the known position is worthless and reckless when it is merely diminished.
Is a superseded number a reason to rewrite the page?
No, and this is the most common real case. A page carrying one stale figure inside otherwise sound structure fails at stage two or three. Replace the number and its link, date the claim in the sentence, and leave the URL and headings alone.
Superseded figures are ordinary. Ahrefs first measured a 34.5% lower clickthrough rate for top-ranking pages on informational keywords across 300,000 keywords, and its December 2025 re-run superseded that with 58%. Any page citing the earlier figure needs one paragraph edited. Treating that as a rewrite fills a content calendar with work that changes nothing, the failure mode described in a visibility gap is not a content brief.
How often should you refresh a page to stay cited?
On no fixed interval. Freshness does show up in the data: Ahrefs' study of about 17 million citations found URLs cited by AI assistants averaged 1,064 days old, against 1,432 for Google's organic results, and a lab study that added dates to passages found LLM rerankers promoted the newer-dated ones, moving the top 10's mean publication year forward by up to 4.78 years. So update the facts, not the date: a new date on unchanged content repairs nothing that actually failed, and Google lists "changing the date of pages to make them seem fresh when the content has not substantially changed" as a warning sign.
The engine moves too. Similarweb recorded the share of ChatGPT's referral traffic going to homepages rising from roughly 26-29% to about 62-63% after the May 2026 update. That measures where ChatGPT sends clicks, not which pages it cites, and it changed while the pages sat untouched.
Name the events that should trigger a refresh: a figure you cite gets superseded, a claim you make stops being true, the answers you track cite a source type you have nothing in, or your logs show the page is no longer fetched. Google states there are no additional requirements or special optimizations for appearing in its AI features, and no engine publishes a refresh interval.
Does publishing something new help on its own?
Not by itself. A new URL that none of your tracked questions reach changes no answer, and it faces the same four stages. If you want to test it, hold the question set fixed and change one variable.
The evidence on rewriting for engines is unflattering. C-SEO Bench, published at NeurIPS 2025, tested conversational-SEO rewrite methods and found most of them largely ineffective and often harmful to citation rank, and the benchmark's authors tested those methods across two tasks and six domains. Phrasing edits are a weak instrument, whatever a refresh checklist implies.
How does Trovance decide refresh versus rewrite?
It keeps the record the four stages are read from, and a person makes the call. Trovance holds the question set still so that diagnosis stays honest. You define the market questions your buyers ask, and the platform runs them on a repeating analysis cycle, preserving each answer run as a snapshot with its full citation set. Because the questions stay fixed, a page dropping out shows up as a change in the record.
Each stage leaves a different fingerprint, and the platform compares those records. Share of answer shows which tracked questions your pages participate in at all. Answer snapshots preserve which URLs carried each answer and what the engine said about you while citing them, which is how a stage-three claims problem separates from a stage-two evidence problem.
The output is a recommended action that names the break the answers show, and no action is a valid result. Your Brand Core holds the claims you are entitled to make and the proof behind each, drafts come from approved claims, and a person reviews everything before it publishes.
Trovance will not promise that a repaired page gets cited, ranked or recommended. Engines are probabilistic, and most of the sources that decide an answer sit outside your domain, so no vendor controls what a model says. What the system does is preserve the evidence and name the break it shows. The next cycle shows you whether the change you shipped altered the answers you were already tracking.
What should you do this week?
Pick the ten pages you would defend in a sales conversation and diagnose those only. Re-run your buyer questions on the schedule you set, record the URLs the answers cite, and sort the pages into the four stages. Most will land in stages one through three, so most teams spend an afternoon on repairs.
A repair is finished when the same question set produces a measurably different answer pattern over enough runs to clear normal variance, never when the document simply looks better. Anything you cannot re-measure is a preference.
The IAB's framework separates presence, prominence, portrayal and persuasion, and those four move for different reasons. A repair that improves portrayal while presence stays flat is still a win, and a blended score would have hidden it.
If you would rather have the question set run for you and the answers preserved as evidence, start a free Trovance trial.
Decide what to fix
A visibility gap is not a content brief - why a gap does not justify a page.
Why AI citations drop - the causes behind a page losing its place.
A relevant page can still miss the proof - the stage-two evidence gap in detail.
How to repair a wrong AI answer about your brand - the stage-three claims repair, mostly off your domain.
Check the mechanics
How AI reads your content - the stages an answer passes through.
AI crawlers don't run your JavaScript - the fetch failure that makes later repairs pointless.
Do headings help AI retrieve long documents - what structure actually contributes to selection.
How long until AI cites you - realistic timing before a repair lands.
