Skip to main content Scroll Top

Why Your AI Visibility Dropped, and Whether It Actually Did

Updated September 2026 · Written and maintained by the Progression Agency strategy team

Roughly a fifth of reported drops turn out to be variance. Rule that out in twenty minutes, then work through six causes on your side and three on theirs.

On this page · 12 sections
  1. First: establish that a drop actually happened
  2. Then: did something change on your side, or theirs?
  3. Read the pattern before you read the cause
  4. The investigation, in order
  5. What you can fix and what you wait out
  6. What not to do while diagnosing
  7. Losing a citation is a different problem from never having one
  8. Everything else we have written on search, AI and getting found
  9. The variance gate, in full
  10. Pattern to cause
  11. Fixable, competable, or wait it out
  12. What not to do

The short answerBefore diagnosing anything, establish that a drop happened: re-run the prompt five times, compare against dated verbatim answers, and confirm wording, platform, region and account state are identical. Variance is structural and a meaningful share of reported drops do not survive that gate. If it is real, roughly sixty per cent of cases trace to a change on your own side — a robots.txt edit, a migration, a CDN or WAF rule, a framework change, a page move or a rewrite — and almost all of those are fixable.

Cause proportions are our own observation across audits, labelled as such rather than presented as research. There is no penalty mechanism in AI answer selection.

Before anything else: is this a real drop?Before anything else: is this a real drop?
In our experience a meaningful share of reported drops do not survive this gate. It costs twenty minutes and it prevents an expensive investigation into nothing.

First: establish that a drop actually happened

Re-run the prompt five times, not once. Compare against stored answers from before. Check the wording is character-for-character identical, and that platform, region and account state match the earlier test.

In our experience a meaningful share of reported AI visibility drops do not survive that gate. Answers vary between runs of the identical prompt, so a single check showing you absent establishes almost nothing.

This costs twenty minutes and it routinely prevents an expensive investigation into something that did not happen. Every other guide to this problem skips straight to causes. We would rather not.

Gate 1 — Five runs. Variance is structural.
Gate 2 — Compare to baseline. No baseline, no drop established.
Gate 3 — Identical wording. One word changes retrieval.
Gate 4 — Same platform and region. Different state, different test.
Gate 5 — Still absent?. Only now diagnose.
Gate 6 — Record it. Dated and verbatim.

Why one check is not enough

Retrieval is influenced by phrasing, timing, region, account state and model version. Appearing in three runs out of five and appearing in none are very different findings that look identical if you only looked once.

Why you need the stored answers

Without a dated record of what the answer said before, you cannot establish that anything changed — only that you are absent now.

Why exact wording matters

Changing a single word changes which sources get retrieved. If the prompt is not identical, you are comparing two different questions.

What counts as the same test

Same platform, same region, same logged-in state. A result from a signed-in session is not comparable to one from a signed-out session.

Then: did something change on your side, or theirs?

Six things on your side are checkable in your own systems today: a robots.txt edit, a migration or redesign, a CDN or WAF rule change, a framework or template change, a page being moved or merged, and content being rewritten.

In what we have diagnosed, roughly sixty per cent of genuine drops trace to a change on the site’s own side — and almost all of those are fixable. That is a more optimistic finding than it sounds.

Did something change on your side or theirs?Did something change on your side or theirs?
The six yeses are checkable in your own systems today. The three ‘part’ rows you cannot control but can often date. The last row is almost never true.
robots.txt edited — Your side. Diff against last month.
Migration or redesign — Your side. The commonest single cause.
CDN or WAF rule — Your side. 403s to bots, invisible in a browser.
Framework change — Your side. Rendering can break silently.
Page moved or merged — Your side. Check redirects.
Content rewritten — Your side. Answer may have left sentence one.
Their side — Model updated. Wait and re-measure.
Their side — Retrieval changed. Wait and re-measure.
Their side — Competitor corroborated. Compete, slowly.
Their side — Category answer shifted. Check the answer shape.
Neither — Variance. Nothing happened.
Neither — Sampling. Your tool ran on a different day.
What actually causes a lost citationWhat actually causes a lost citation
Our own observation across audits, labelled as such rather than presented as research. Sixty per cent is your own side and fixable; eighteen per cent was never a drop.

Read the pattern before you read the cause

The shape of the loss narrows it considerably. Sitewide and across every platform points hard at access or rendering. One platform sitewide points at that platform’s crawler. One page across every platform points at that page.

One page on one platform, with no other movement, is usually variance — and that is the combination that most often triggers investigations that find nothing.

Locating the cause from the patternLocating the cause from the pattern
Sitewide and every-platform points hard at access or rendering. One page on every platform points at that page. One platform sitewide points at that platform’s crawler.
Sitewide, all platforms — Access or rendering. Check robots.txt and scriptless fetch.
One platform, sitewide — That crawler. Check that agent specifically.
One page, all platforms — That page. Did it move or change?.
One page, one platform — Likely variance. Re-run before investigating.
Gradual across months — Corroboration. Competitors gained, you did not.
Sudden and total — Almost always your side. Something shipped.

The investigation, in order

Everything before hour two is free and rules out the most common causes. Re-run the prompt. Diff your robots.txt. Fetch the page with JavaScript disabled and compare the word count to what it was. Request it with the AI user agents and confirm a 200.

Most investigations end inside those four steps, usually on the robots.txt diff or the scriptless fetch. A migration or a security plugin update is the single most common culprit we encounter.

The investigation, in orderThe investigation, in order
Everything before hour two is free and rules out the most common causes. Most investigations end there.

Diff your robots.txt

Against version control or a web archive copy. A Disallow added during a migration is the cleanest explanation there is, and the easiest to miss.

Compare the scriptless fetch

Not just whether it returns content, but whether it returns as much as before. A framework upgrade can move content behind JavaScript without anyone noticing.

Check status codes by user agent

A WAF rule tightened last month may now return 403 to unrecognised agents while serving browsers perfectly.

Read the logs for frequency change

Crawler hits dropping off is a dated signal. It tells you roughly when, which usually tells you what.

Check whether the page itself changed

Moved without a redirect, merged into another page, or rewritten so the answer is no longer in the first sentence of its section.

What you can fix and what you wait out

Six causes are fixable by you: a robots.txt regression, a WAF blocking AI agents, rendering broken by a framework change, a page removed without a redirect, an answer displaced during a rewrite, and entity data that has drifted out of alignment.

Two you cannot: a model version change and a platform changing its retrieval. For those the honest advice is to re-measure in thirty days rather than react. One sits in between — a competitor gaining corroboration — which you can compete with, over months.

What you can fix, and what you wait outWhat you can fix, and what you wait out
Six fixable, one you can compete with over months, two you can only wait out. Knowing which category you are in is the entire value of the investigation.
robots.txt regression — Fix now. Free, minutes.
WAF blocking agents — Fix now. Low cost.
Broken rendering — Fix now. Engineering, but known.
Missing redirect — Fix now. Free.
Answer buried in a rewrite — Fix now. Low cost.
Entity drift — Fix now. Weeks, systematic.
Wait out — Model version change. 30 days, then re-measure.
Wait out — Retrieval change. Same.
Compete — Competitor corroboration. Months, not weeks.
Do not — Panic-publish content. Volume is not the lever.
Do not — Rewrite everything. Diagnose first.
Do not — Assume a penalty. No such mechanism exists here.

What not to do while diagnosing

Do not panic-publish content. Do not rewrite pages that were working. And do not assume a penalty, because there is no penalty mechanism in AI answer selection in the way there is in search.

Absence is an eligibility or selection outcome, not a sanction. Reacting as though it were a punishment leads to sitewide rewrites that destroy the structure that was previously earning the citation.

Diagnosing a drop, in numbersDiagnosing a drop, in numbers
The zero is the one that matters. Without stored prior answers there is no way to establish that anything changed, let alone why.

Losing a citation is a different problem from never having one

If you previously had it, your own recent changes are the first suspect and variance is the second. If you never had it, eligibility is the question — whether the machine could reach and read you at all.

The two investigations start in different places, which is why we keep them on separate pages. If you have never been cited, the diagnostic for that is a different sequence entirely.

Lost a citation versus never had oneLost a citation versus never had one
If you previously had the citation, your own recent changes are the first suspect. If you never had it, eligibility is — which is a different page on this site.

Cannot tell whether it dropped or never registered?

An AI visibility audit establishes a dated baseline with verbatim answers, checks access and rendering, and tells you whether you are looking at a regression, variance, or a citation you never actually had.

/ai-visibility-audit

Everything else we have written on search, AI and getting found

AI, AEO and what is changing

Websites and design

Choosing and working with an agency

Social, content and brand

By industry and by situation

Want this done for your site?We build and maintain the search, content and paid programmes described on this page.

Get a free proposal

The variance gate, in full

Before diagnosing a drop, establish whether anything actually dropped — this is the test for that.

Rule this out before diagnosing anything
CheckWhat to doIf it failsTime
Multiple runsRe-run the prompt five times in one sittingYou observed variance, not a drop10 min
Baseline existsCompare against stored verbatim answersYou cannot establish a drop at all5 min
Wording identicalCompare exact prompt stringsYou compared two different questions2 min
Same platformConfirm which assistant, which modeNot a comparable test1 min
Same regionConfirm location settings matchAnswers are location-sensitive1 min
Same account stateSigned in versus signed outNot a comparable test1 min

Pattern to cause

Work from the shape of what you are seeing to the cause that produces that shape.

What the shape of the loss tells you
PatternMost likely causeFirst checkFixable?
Sitewide, every platformAccess or rendering brokerobots.txt diff, then scriptless fetchYes
Sitewide, one platformThat platform’s crawler blockedThat agent in robots.txt and WAFYes
One page, every platformThat page changedDid it move, merge or get rewritten?Yes
One page, one platformUsually varianceRe-run five timesNothing to fix
Gradual over monthsCompetitor corroboration improvedSearch for their recent mentionsSlowly
Sudden and totalSomething shipped on your sideDeployment log for that weekYes

Fixable, competable, or wait it out

Three categories of cause, and the very different responses each one warrants.

Sorting causes by what you can actually do
CauseCategoryActionTime to resolve
robots.txt regressionFixableRestore the allow rulesMinutes
WAF or CDN blocking AI agentsFixableAllowlist the user agentsHours
Rendering broken by a changeFixableRestore server-side renderingDays to weeks
Page moved without redirectFixableAdd the redirectMinutes
Answer displaced in a rewriteFixableMove it back to sentence oneMinutes per page
Entity data driftedFixableRe-align every sourceWeeks
Competitor gained corroborationCompetableEarn your ownMonths
Model version changedWaitRe-measure in 30 daysUnknown
Platform changed retrievalWaitRe-measure in 30 daysUnknown

What not to do

The reactions that make a drop worse, which are also the most common ones.

Common reactions that make it worse
ReactionWhy it is wrongDo this instead
Publish a burst of new contentVolume is not the lever and it delays diagnosisDiff your own recent changes
Rewrite pages that were workingYou may remove the structure that earned the citationChange nothing until you have a cause
Assume a penaltyNo penalty mechanism exists in answer selectionTreat it as eligibility or selection
React to one checkVariance is structuralRe-run five times first
Switch platforms to testDifferent surface, different resultHold platform constant
Blame the agency immediately60% of causes are site-side changesCheck the deployment log

Getting found in search

AI, AEO and what is changing

Paid media and lead generation

Websites and design

Choosing and working with an agency

Social, content and brand

By industry and by situation

Frequently asked questions

Why did my AI visibility drop?
First establish that it did. Re-run the prompt five times and compare against stored answers — variance is structural and a single check proves very little. If it is a real drop, roughly sixty per cent trace to a change on your own side: a robots.txt edit, a migration, a CDN or WAF rule, a framework change, a page move, or a rewrite.
How do I know if it is a real drop or just variance?
Re-run the prompt five times in one sitting, compare against dated verbatim answers from before, and confirm the wording, platform, region and account state are identical. A meaningful share of reported drops do not survive that gate.
What is the most common cause of a lost AI citation?
A change on your own site — most often a migration or redesign that altered robots.txt, rendering, or the page itself. This is good news, because it is the fixable category.
ChatGPT used to mention us and now it does not. What changed?
Check your own side first: diff robots.txt against last month, fetch the page with JavaScript disabled and compare the word count, and confirm AI user agents still get a 200 rather than a 403.
Could a model update have caused it?
It could, and that is one of the two causes you cannot fix. If everything on your side checks out, the honest advice is to re-measure in thirty days rather than react.
Is there a penalty in AI answers, like a Google penalty?
No. There is no penalty mechanism in answer selection. Absence is an eligibility or selection outcome, and treating it as a sanction leads to sitewide rewrites that often remove the structure that was earning the citation.
How long should I wait before investigating?
Do the twenty-minute variance gate immediately. If it survives that, investigate straight away — the free checks take under two hours and rule out the most common causes.
What does a sitewide drop across every platform mean?
Almost always access or rendering. Check robots.txt first, then whether a scriptless fetch still returns your content, then whether your CDN or WAF began returning 403 to AI user agents.
What if only one platform stopped citing me?
That points at that platform’s specific crawler rather than at your content. Check that agent in robots.txt, and whether your WAF started blocking it.
What if only one page lost its citation?
Check whether that page moved, merged, or was rewritten — particularly whether the answer is still in the first sentence of the section whose heading promises it.
Can I diagnose a drop without a baseline?
Not really. Without dated verbatim answers from before, you cannot establish that anything changed, only that you are absent now. That is the strongest argument for recording one before you need it.
A competitor is now cited instead of us. What happened?
Either something broke on your side, or they gained corroboration. Check your own changes first — it is faster and more likely — then look for recent third-party mentions of them.
Should I publish new content to recover?
No, and not yet. Volume is not the lever and publishing delays diagnosis. Find the cause first; most of the time it is a configuration regression that takes minutes to fix.
How do I check if a migration caused this?
Diff robots.txt against an archived copy, compare a scriptless fetch word count against what the page used to return, check redirects for any moved URLs, and read your deployment log for the week the drop began.
What does a gradual decline over months indicate?
Usually corroboration rather than a technical fault. Competitors accumulating independent mentions while you did not is slow, undramatic, and the hardest category to reverse quickly.
How often should I re-measure after a platform change?
Thirty days. Reacting inside that window mostly produces changes you will not be able to attribute afterwards.
Does a Google ranking drop explain an AI citation drop?
Not directly, though a shared cause can produce both — a migration that broke indexation would affect an index-led AI surface too. They are different surfaces and should be diagnosed separately.
What should I record so this is easier next time?
A dated prompt set with verbatim answers, monthly. It converts ‘I think we used to be cited’ into evidence, and it is an hour a month.
Can a tool tell me why I dropped?
It can tell you that you are absent. It cannot tell you why a model chose someone else, because no provider exposes that reasoning. The diagnosis comes from your own site and change history.
How much of this can I do myself?
All of the free checks — the variance gate, the robots.txt diff, the scriptless fetch and the user-agent test. That is under two hours and it resolves most cases.
What if everything checks out and I am still absent?
Then you are likely looking at a platform or model change, or a competitor who improved. Re-measure in thirty days and, if it persists, treat it as a corroboration problem rather than a technical one.
Is losing a citation the same problem as never having one?
No. If you had it, your own recent changes are the first suspect. If you never had it, the question is eligibility — whether machines can reach and read you at all — which is a different diagnostic sequence.

Want this done for your site?We build and maintain the search, content and paid programmes described on this page.

Get a free proposal

Get a free marketing proposal

Tell us what you are trying to grow and we will come back with a plan, not a pitch deck. Same-day reply on weekdays.

Our website uses cookies from third party services to improve your browsing experience. Read more about this and how you can control cookies by clicking "Privacy Preferences".
Privacy Preferences
When you visit our website, it may store information through your browser from specific services, usually in form of cookies. Here you can change your privacy preferences. Please note that blocking some types of cookies may impact your experience on our website and the services we offer.
Privacy Policy
You have read and agreed to our privacy policy
Required
Contact Us