When an AI answer changes, the website is only one suspect. The cited source may have moved, the engine may have sampled differently, or a competitor may have become easier to describe.
A heating company near Vannes once gave me the kind of sentence that makes a measurement ledger useful: “Nothing changed on the site, but Copilot stopped naming us.” In the ledger, the drop looked clean at first. The business had appeared in five of eight local maintenance prompts in the previous run. Three weeks later, it appeared in one. The service pages were still online. The Google Business Profile entries had not been edited. The branch pages had the same titles, the same photos, even the same awkward paragraph about boiler servicing that nobody had touched for years.
This is a composite scenario, assembled from several regional service audits, but the odd detail is familiar: the company was still visible for emergency repair prompts, including one town where it had no technician most days, while it disappeared from the more valuable commercial maintenance prompts. That is the sort of thing a normal SEO dashboard rarely explains. Nothing “happened” in the place the team was watching. Something happened in the answer path.
The website is not the whole evidence field
When a team sees a drop in AI visibility, the first instinct is to inspect the last website change. Someone edited a heading. Someone removed a paragraph. Someone changed a location page. That reflex comes from ordinary search work, where the site often feels like the main instrument. I understand it. I have spent enough time with title tags, local pages and crawl reports to know the habit.
AI answers make the field wider. A generated answer may use the business site, but it may also lean on directories, review snippets, partner pages, map results, trade articles, old PDFs, comparison pages, vendor listings and copied descriptions sitting in places no one in the company has opened for years. The answer is stitched from what the engine can retrieve, trust, summarize and fit into the user’s prompt.
That means a fall in visibility can arrive without a site change because the engine’s usable evidence changed somewhere else. A directory page may have reshuffled. A competitor may have gained a clearer third-party mention. A cited source may have been updated in a way that changes category wording. The engine may have altered how it blends web retrieval with model memory. Sometimes the same prompt is simply sampled differently enough to change a borderline answer.
AI visibility drift is a change in repeated answer behavior, because the evidence path has shifted even when the company’s own pages have not.
I use that definition because it stops the panic from narrowing too early. The cause might be on the site, yes. It might also be upstream, downstream or beside it. The ledger has to preserve that uncertainty long enough to find the moving part.
A drop is not one missing answer
The most common false alarm is one prompt returning one bad answer. A manager writes “best heating maintenance company near Rennes” into ChatGPT, sees three competitors and no company name, then sends the screenshot around like a fire alarm. The screenshot may be useful as a clue. It is not a measurement.
A drop needs a before and after sample. It needs the same prompt set, or a deliberately revised one with the revision marked. It needs dates, engines, languages and location intent. It needs enough repetition to distinguish a pattern from ordinary answer variation. That does not mean hundreds of runs for every small business. It means refusing to treat one answer as the whole room.
In my prompt ledgers, I separate four fields before I read the loss as meaningful: presence, cited source, answer position and description quality. Presence asks whether the business is named. Citation asks which page, if any, is used as the evidence. Position asks whether the name appears high enough to be commercially visible. Description quality asks whether the answer still says the correct thing. A business can lose position but keep presence. It can keep citation but lose description accuracy. It can be named more often while being described worse.
For the regional heating network, the first sign was not absence everywhere. It was a shape change. The company remained present in emergency prompts and broad “plumber near me” prompts. It weakened in maintenance prompts, especially in French. In English, it was barely present before and barely present after, so there was no meaningful drop there. The loss belonged to one part of the buyer path.
That matters because a broad diagnosis would have produced broad waste. Rewrite everything. Add more service pages. Publish more articles. Ask for more reviews. All possible, all expensive, and most of them premature. The ledger said something narrower: maintenance evidence had become less retrievable or less convincing than emergency evidence.
Source shifts are quieter than site edits
A source shift is the least dramatic kind of visibility drop, which is why teams miss it. Nobody in the company changes anything, but an answer engine starts citing a different page for the same category. The old source might have been the company’s own branch page. The new source might be a directory listing, a review page, a competitor comparison, or a general local guide.
In the heating network scenario, the answer had previously cited a branch page for “commercial boiler maintenance in [city].” Later it cited a general directory page that grouped the company with emergency plumbers. The company was still named sometimes, but the surrounding description changed. The answer leaned on the emergency category because the source it used was built that way. The website did not lose the fact. The cited source did.
This is where ordinary monitoring has to be rather plain. I want the exact prompt, the engine, the language, the location intent, the names returned, the cited source and the description error in separate fields. I do not want a single score too early, because the score hides the movement. A falling number tells me almost nothing if I cannot see whether the cited source changed.
There is another roughness here. Sometimes the source shift is partial. Perplexity may continue citing the company’s own page, while Copilot begins citing a directory. Google AI Overviews may show no clean citation at all, or may blend sources in a way that makes attribution awkward. ChatGPT may cite nothing in one run and cite a page in another. This is why each engine keeps its own column. A combined average can make three different problems look like one polite decline.
The useful question is not “Why did AI stop liking us?” That sentence is too human, too tidy. The better question is: “Which repeated answer path changed, and what source now feeds it?”
Model shifts look like bad weather until they repeat
There is also the harder case: the evidence does not visibly change, but the answer behavior does. I do not pretend this is always easy to explain. Public answer systems adjust retrieval, ranking, summarization, citation display and safety rules. The outside observer sees the surface. The machinery is mostly behind the wall.
Still, the surface leaves tracks. A model or retrieval shift often appears across several prompts at once. It may affect categories more than locations, or English more than French, or broad buyer questions more than specific brand questions. It may change the style of answers: fewer small local firms, more directories; fewer exact service descriptions, more generic category text; fewer named sources, more blended summaries.
In most cases, the clue is not one company’s disappearance. It is a change in the answer set. Three competitors move. Citations concentrate around a different kind of source. The engine begins preferring pages with fresher structured facts, or pages with stronger third-party descriptions, or pages that are easier to summarize. I phrase that carefully: the data suggests these movements; it does not let us see the whole internal cause.
For a French SMB, the practical answer is not to chase the model. It is to run a stable sample often enough to notice when the pattern bends. If a business measures once a year, a drop looks like a mystery. If it measures monthly, and keeps the prompt set stable, the loss has a date, a shape and a smaller group of suspects.
A monitoring rhythm is not there because the dashboard is pretty. It is there because memory is bad at this. People remember the painful screenshot and forget the three neutral ones. They remember the competitor that appeared and forget that it already appeared last month. A ledger is a dull object, but it is loyal.
The wrong repair can make the next reading worse
When visibility drops without a site change, the worst response is often an immediate rewrite. Not because rewriting is useless. It can be the right repair. But rewriting before the source path is understood is like repainting a sign because a delivery driver used the wrong map.
If the drop comes from a weak third-party source, the better repair may be to strengthen the company’s own page and also correct or outcompete the external source. If the engine is citing a directory that frames the business as emergency-only, the company needs clearer maintenance evidence in places the engine can retrieve. That may include its branch page, service descriptions, case evidence, public profiles and category wording. The exact mix depends on what the ledger shows.
If the drop is mostly a language split, the repair is different. French pages may carry the commercial maintenance proof, while English vendor or directory pages describe only emergency services. A bilingual company can look like two different entities to an answer engine. Testing French and English separately prevents the English evidence path from being mistaken for the French one.
If the drop is a position loss rather than disappearance, the repair changes again. The company may still be named, but lower than competitors with stronger cited sources. That points toward citation share and source quality, not simple presence. This is where a sibling measurement on competitor citation share becomes useful, though it is a separate question from the drop itself.
The discipline is to keep the failure type intact. Absence, weak citation, lower position and bad description are different failures. They sometimes arrive together, but they do not ask for the same tool.
How I monitor the fall without inventing certainty
My monitoring sheet for a visibility drop is intentionally unglamorous. I keep the old prompt set frozen for comparison. If we add new prompts, they sit in a separate group so they do not corrupt the baseline. I run French and English separately. I mark location intent plainly: national, regional, city, nearby town, service area. I record whether the company appears, where it appears, who else appears, what source is cited and what the answer says wrong.
Then I read the movement in layers. First, did presence change? Second, did position change? Third, did cited sources change? Fourth, did description accuracy change? Fifth, did the competitor set change? Only after that do I form a judgment about likely causes. Even then I prefer “most likely” to “proved,” unless the evidence is clean.
For the heating network, the most useful finding was modest: the loss was concentrated in French maintenance prompts around branch-level locations, and the cited sources had moved toward emergency plumbing directories. That was enough to stop a general rewrite plan. The repair started with clearer maintenance evidence on branch pages and a correction pass on weak external profiles. Retesting came after, not before.
A visibility drop without a site change is not a paradox. It is a reminder that AI answers live in a wider evidence field than the website. The business page may be steady while the path toward it becomes muddy.
The Measurement Note — Signal: visibility falls in one prompt group while the site stays unchanged. Distortion: blaming the last page edit because it is the only visible event. Ledger: record old and new answer presence, position, cited source, language and description error for the same prompts. Next Test: rerun last month’s prompt set unchanged before writing any new copy, then mark where the source path moved.