Your Client's AI Mentions Dropped. Here Is How to Diagnose It Before the Call

The message arrives on a Tuesday. "We ran a few searches and we're not showing up anymore. What happened?" Last month's report showed steady gains. Now you have a client who tested three prompts by hand, found nothing, and drew a conclusion.

Most agencies handle this badly in the same two ways: they either promise to investigate and go quiet for a week, or they immediately start changing content, which destroys the only baseline they had.

There is a faster path. Five causes account for essentially every reported drop, they can be ruled out in a fixed order, and four of the five are not caused by anything you did. Running that sequence takes about 45 minutes and turns a defensive call into a demonstration of competence.

Rule zero: confirm something actually dropped

Before diagnosing, establish that there is a phenomenon. Client-reported drops frequently come from three prompts typed once each into a logged-in ChatGPT session on a phone. That is not a measurement, and the research on why is unusually clear.

A 2026 variance decomposition study analyzing 12,933 LLM responses across 20 brands, 8 languages, and 3 models found that brand identity accounts for roughly 1.5 percent of total variance in an answer, and concluded that a single AI answer carries almost no brand-discriminating signal. Read that again in operational terms: one run tells you close to nothing about whether a brand is visible. Three runs tell you slightly more than nothing.

The same study found within-prompt resampling accounts for around 35 percent of variance on its stability subset, and that a repeat past the fifth reduces relative-error variance by only 0.0003. Five runs per prompt per engine is the point where you stop buying meaningful precision. That is your minimum bar for calling anything a drop.

The first thing to say on the call is not "we found the problem." It is "here is what we measured and how, here is what the client measured, and here is why the two differ." That reframes the conversation from blame to method, and it holds up because it is true.

Notably, the same research found that adding languages and models reduces measurement error about 15 times more effectively per unit of budget than adding more repeats of the same prompt. If your monitoring runs one engine, your noise problem is structural and no amount of re-running fixes it.

The five causes, in the order to rule them out

Work top to bottom. Each step is cheaper than the one below it, and stopping early saves you from changing content that was never the problem.

The order matters because causes 1 through 3 are external and cause 5 is the only one you can fix directly. Agencies that start at 5 spend a week auditing a site that was fine.

Cause 2: the whole platform moved

This is the cause most agencies skip and most often should check. Engines change citation behavior at the platform level, and when they do every brand in the category moves at once.

seoClarity's tracking of ChatGPT citation volumes across five markets between February and May 2026 documents exactly this. Citation volumes fell 86 to 94 percent across all tracked markets, with an initial inflection on March 8 and a severe collapse around April 19 where the US, UK, and Germany each saw volumes fall by over 80 percent. The US zero-citation rate, meaning answers that cite nothing at all, doubled from 28 percent to 48 percent in March. Germany reached 85 percent by April. Volumes then rebounded in May toward pre-March levels.

The mechanism there was not brands losing authority. It was ChatGPT answering more questions without citing any source, which looks identical to a visibility drop in any tool that counts citations.

The tell is portfolio-wide correlation. If six of your eight clients dropped on ChatGPT in the same seven-day window while their Perplexity numbers held, you have a platform event, not eight simultaneous content failures. Agencies have a diagnostic advantage here that in-house teams do not, and it is worth saying so on the call.

Two things follow. First, never report citation counts without also reporting the zero-citation rate for that engine, or you will attribute an engine's behavior change to your own work in both directions. Second, if a platform event is in progress, the correct action is to wait and document, not to rewrite content. The May rebound in that dataset would have made any panic rewrite look like a brilliant fix.

Cause 3: one engine changed how it retrieves

If the drop is isolated to a single engine while others hold, you are looking at a retrieval or model change on that engine rather than anything about the brand.

The practical test is a same-page comparison. Take a page that was cited on the affected engine last month and check whether it is still cited on the others. If Perplexity and Gemini still quote it and ChatGPT no longer does, the page did not get worse. Something in one retrieval stack reweighted.

This happens because each engine sources differently, weights recency and corroboration differently, and generates different synthetic sub-queries from the same user question. Our breakdown of why one AI engine cites you while another ignores the same page covers the specific mechanics per engine, which is what you need in order to explain the divergence rather than just observe it.

Model version splits are a related and underdiagnosed variant. When an engine runs multiple model versions concurrently, different users hitting different versions can get materially different brand lists for the same question. A client on one version and your monitoring on another will disagree, and neither is wrong.

Cause 4: you did not fall, someone else rose

Citation rate and share of voice move independently, and confusing them produces the wrong diagnosis.

Your citation rate can hold at 40 percent while your share of voice falls, because a competitor entered the answer set and the answer now names four brands where it named two. Nothing about your position degraded. The category got more crowded.

Check for a new name in the answer text. If one appears, ask what evidence the engine found for them, published when, and on which domain. That is usually recent third-party coverage or a review-site placement rather than anything on their own site.

This distinction changes the recommendation entirely. A citation rate drop points at your content. A share of voice drop with flat citation rate points at earned media and corroboration. Reporting only one of the two numbers guarantees you will sometimes prescribe the wrong fix, which is why the metric hierarchy in our guide to which AEO metrics actually predict pipeline separates them.

Cause 5: something on the client's side changed

Last, and least often, the client changed something. This is the cause worth checking most thoroughly when the drop appears across every engine simultaneously, because a genuinely cross-engine drop is rarely coincidental.

Run this list in order:

1. Crawler access. Did a robots.txt change, a WAF rule, a bot-protection upgrade, or a CDN configuration start blocking GPTBot, OAI-SearchBot, ClaudeBot, or PerplexityBot? This is the single most common real cause and it is often done by an infrastructure team that never heard of the AEO program.
2. URL changes. Did a site migration, a slug change, or a category restructure break the URLs engines had indexed? Redirects preserve link equity for Google far better than they preserve citation continuity for AI engines.
3. Removed or consolidated pages. Content pruning is standard SEO hygiene and routinely deletes the exact pages engines were quoting.
4. Rendering changes. A move to client-side rendering can make content invisible to crawlers that do not execute scripts, with no visible change to human users.
5. Answer block edits. Did anyone rewrite the direct-answer paragraph at the top of a cited page? Engines quote specific passages, and editing the quoted passage removes the thing being quoted.

Server logs settle most of this in minutes. If AI bot hits to the affected URLs stopped on a specific date, you have both the cause and the date, and the conversation becomes a fix rather than a theory.

If discovery turns out to be the issue, the mechanical repairs are cheap. Restoring crawler access, republishing a clean machine-readable content index with a tool like our free llms.txt generator, and serving content in a form engines can ingest without executing scripts, which is what our AI Feed Engine handles across a portfolio, will typically restore crawl within one to two weeks.

The 45-minute run sheet

Do this before you reply to the email, not after.

Minutes 0 to 10. Pull the tracked prompt set results for the affected client for this month and last, five runs per prompt per engine. Confirm whether the tracked data shows the drop the client reported. Note the zero-citation rate per engine.

Minutes 10 to 20. Check your other clients on the same engine for the same window. Portfolio-wide movement means platform event, and you can stop here for the causal question.

Minutes 20 to 30. Split the result by engine. Isolated to one engine means retrieval change. Present on all engines means look at the client side.

Minutes 30 to 40. Read the actual answer text, not just the scores. Who is named now who was not named before? Did the answer stop citing anyone at all? Those two findings point at completely different causes.

Minutes 40 to 45. Check server logs for AI bot hits to the affected URLs, and check robots.txt against last month's version.

You will finish with a cause, a date, and evidence. That is a materially different call from "we're looking into it."

This runs in 45 minutes rather than three days only because the baseline already exists. Continuous tracking across engines is what makes a drop diagnosable at all, which is the point of the monitoring layer in how OnlyAEO works. Without a prior month of five-run data across engines, you are reduced to guessing.

What to say on the call

Three rules keep the conversation productive.

Lead with the method, not the verdict. "You tested three prompts once each. We test 40 prompts five times across four engines. Here is what our data shows for the prompts you ran." This is not defensive if you follow it immediately with the finding.

Name the cause plainly, including when it is your side. If a crawler got blocked during a security update, say that, with the date and the log evidence. Clients forgive diagnosed problems. They do not forgive vagueness.

Say what happens next and when. Platform events call for documenting and waiting, with a specific re-check date. Crawler blocks call for a fix and a two-week re-crawl window. Competitor displacement calls for an earned media plan with a longer horizon. Each has a different timeline and saying the wrong one sets up the next difficult call.

Volatility itself is worth pre-framing in every engagement. The causes of AI search volatility include model updates, retrieval timing, personalization, and competitive changes, none of which are under your control. A client told in month one that month-to-month movement of several points is normal reacts very differently in month four than a client who was promised a smooth line.

Make the next drop cheaper

Four changes turn drop diagnosis from a fire drill into a lookup.

• Track five runs per prompt across at least three engines. Anything less and you cannot distinguish noise from signal, and the variance research says added engines beat added repeats by a wide margin.
• Report the zero-citation rate alongside citation counts. It separates engine behavior from brand performance.
• Snapshot robots.txt and the answer text monthly. Both are free to store and impossible to reconstruct later.
• Put a volatility band in the monthly report. The structure in our guide to what to put in a monthly AEO report for agency clients includes how to present a down month without losing the room.

Programs that survive their first bad month are usually the ones that predicted the bad month out loud. Our FastTrackr AI case study shows what the real curve looks like across months, including the flat and down stretches that a projection chart never has.