Why Your AI Visibility Dropped

The report lands and the number has fallen. Appearances down from 19 to 11. Perplexity halved. The mention in AI Overviews gone entirely. Nothing was deployed last month. Somebody wants an explanation by Thursday.

Before anything else: there is a good chance nothing dropped. AI answers are probabilistic, and a fall of that size can be ordinary variance in a small sample. The first job is not to find the cause. It is to establish whether there is one.

First, check it actually dropped

Two tests separate signal from noise, and both take minutes.

Is the fall spread or concentrated? A drop scattered thinly across many prompts and engines, a mention lost here, another there, is almost always variance. A drop concentrated in one place has a cause, and you can usually name it within the hour. Two shapes to look for: every prompt on one engine gone, or every appearance that used to cite one particular page gone. The first points at the engine, the second at the page.

Did the instrument change? If the prompt set was edited, a prompt added or reworded, the engine mix adjusted, the number of runs per prompt reduced, or the tool itself swapped, the comparison is invalid and there is nothing to investigate. This is the most common explanation of all and the least often considered. It is also the argument for freezing the prompt set, as we set out in what an AI Visibility Score is.

A 30-minute triage

If the fall survives both tests, work through this in order. It is sequenced by how cheap the check is and how often it turns out to be the answer, not by how serious the cause would be.

1. Fetch the cited pages as an AI user agent. Not as yourself in a browser. Request each previously cited URL with the user agent string of GPTBot, PerplexityBot and Google-Extended, and confirm you get a 200 with the content present in the raw HTML, not just on the rendered page. Those are different things, and the difference is the whole finding.
2. Check those URLs still exist at the same address. A slug change with a redirect in place is still a change of address, and citations do not always follow.
3. Check robots.txt and any AI-specific directives. Including whether something was added to it that nobody announced.
4. Check the firewall. Covered below, and more often the answer than anyone expects.
5. Read the answers that replaced yours. Who is named now? A new competitor source is a different problem from a generic answer naming nobody.
6. Check the third-party sources that used to carry you. Directories, review platforms, trade press.

Six checks. In our experience one of the first four resolves it most of the time.

Causes that are your side

A security change started blocking AI crawlers

This is the one we would check first on any site that has had security work done recently, and it is badly under-diagnosed. Firewall rules, bot-protection settings and managed rulesets frequently filter on user agent, and a ruleset tightened to stop scrapers will stop GPTBot and PerplexityBot with equal efficiency. The site looks perfect to you, because your browser sends a browser user agent.

The giveaway is a drop that is sharp, total and dated: visibility does not decay, it stops. If the fall lines up with a deployment, a plugin update, a CDN change or a hardening exercise, start here. Test by requesting the same URL twice, once with a browser user agent and once with a bot user agent, and compare the status codes. The difference is the answer.

The cited page changed or disappeared

Citations attach to pages, not to brands. If the page that carried your appearances was consolidated, re-slugged, reworked, set to noindex or quietly unpublished during a tidy-up, the appearances go with it. A redirect preserves link equity far more reliably than it preserves a citation.

This is the strongest practical argument for breadth. If your appearances draw on one page you have a single point of failure, which is why source breadth belongs in the report alongside the headline count.

Rendering changed

Content that moved behind JavaScript, a consent wall, a tab, an accordion or a lazy-loaded component can vanish from the raw HTML while looking identical to a human. Check what is actually in the response body rather than what is on the screen: see what AI reads on your website.

A rebrand or merger scrambled the entity

Renaming, merging or restructuring breaks the association between a name and a body of evidence, and engines become cautious about naming an organisation they cannot resolve. Expect a dip for a period after any identity change, and work through the rebrand and merger SEO checklist rather than waiting it out.

The content went stale against a fresher source

Where a question has a time-sensitive answer, rates, thresholds, regulations, engines prefer the source that looks current. A competitor publishing an updated version of your guide can displace you without doing anything else. The remedy is a genuine update with a real change of substance, not a touched modified date.

Causes that are not your fault

Some of these are worth reporting precisely because no action follows.

The engine changed how it cites. Platforms adjust how many sources they name, whether they link, and how much they rely on their own summaries. A change that reduces citations across the board will reduce yours, and every competitor’s, with no signal to you.

The answer stopped naming anyone. Sometimes an engine moves from naming three firms to giving a generic explanation. You did not lose to a competitor; the prompt stopped being a branded answer. Check who replaced you before assuming it was someone.

A third-party source dropped you. If a directory delisted you, a publication archived the article that quoted you, or a review platform shut down, the citations resting on it go too. This is the risk inherent in visibility you do not own.

The index was resampled. Where the figure comes from a standing index of AI responses, the composition of that index shifts. A change in what the provider collected is a change in your number that has nothing to do with your site.

Say so plainly when this is the explanation. A report that attributes every movement to the agency’s work is less credible than one that occasionally says the cause was external and no action is recommended.

What not to do

Do not rewrite the page that lost its citation. Not until you know why. If the cause was a blocked user agent, rewriting a page nobody can fetch is a week spent on nothing.

Do not change the prompt set in response. It is tempting to add prompts you are confident about. It destroys the only baseline you have.

Do not switch tools mid-investigation. A second tool will give a different number because it uses a different method, and you will not be able to tell which part of the difference was the drop.

Do not act on a single period. Unless the fall is sharp, total and traceable to a date, wait for the second measurement. Most apparent drops are gone by then, and the ones that are not have a cause you can name.

FAQ