Someone senior typed your category into ChatGPT, got a competitor's name back, screenshotted it, and sent it to you with a one-line message. You are now expected to explain it, probably today, and possibly with a plan attached.
This page is the sequence for producing that explanation without guessing. It ranks the four realistic causes by how often they turn out to be the answer, and gives each one a test you can run before anyone approves a budget. Two of the four require no remediation at all. That is not a comforting framing, it is the statistically likely one.
What You're Seeing, And What It Does Not Yet Prove
One answer, from one surface, on one day, from one account, in response to one phrasing.
That observation is real. Here is the complete list of what it establishes: on that occasion, that system produced that text. It does not establish that the competitor is "winning," that your visibility has fallen, that anything changed, or that the same prompt would produce the same answer if you ran it again five minutes later. Assistants are non-deterministic; repeated identical prompts routinely produce different named brands. So the first thing to say out loud, in whatever channel the screenshot arrived in, is a version of this: we have one sample, we are running the measurement now, and we will have a defensible answer in a few days. That sentence buys you the time to do this properly and costs you nothing, because it is true.
What you must not do is act on the screenshot. Every subsequent step in this page exists to stop you spending money on a problem that has not yet been shown to exist.
Cause 1: You Sampled Once
This is the most common cause by a wide margin, and it is the cheapest to rule in or out.
Because these systems are non-deterministic, any single answer is one draw from a distribution. If your brand appears in, say, a bit less than half of answers to a given prompt, then a single run will show a competitor instead of you more often than not, and nothing whatsoever has changed. A single unfavourable answer is entirely consistent with strong visibility.
The test. Take the exact prompt from the screenshot. Run it repeatedly through the official API for that surface, in fresh sessions, over a short window. Record for each run: whether your brand was named, which competitors were named, whether sources were returned, the model version if exposed, and the timestamp. Then look at the proportion of runs naming your brand, not at any individual run.
How to read the result. If your brand appears in a meaningful share of runs, the screenshot was a draw from an existing distribution and there is nothing to diagnose. If your brand appears in a small share or none, you have a real signal worth carrying to Cause 2. How many runs you need before that judgement is trustworthy is a sample-size question with an actual answer, covered in how many runs you need. Guessing at ten because ten is a round number is how people talk themselves into false conclusions in both directions.
What not to do yet. Do not compare this run against a number from before, because you almost certainly do not have a comparable one. That is the next problem.
Cause 2: Your Prompt Phrasing Selected Against You
The second most common cause is that the prompt was never a fair test.
Prompts are not neutral. A question phrased around a feature you do not lead on, a use case you do not serve, a price point you sit above, or a geography you do not operate in will reliably surface companies that fit it better. That is the system working correctly. The screenshot then shows a real answer to a question your buyers may never ask.
The test. Write out the prompt from the screenshot and answer three questions about it in writing.
- Does this describe a buyer you actually want, at a stage where they would plausibly be asking an assistant?
- Does it use language your buyers use, or language your competitor's marketing uses?
- Does it contain an implicit qualifier — cheapest, simplest, for enterprise, for a specific region, open source — that excludes you by definition?
Then build a small, deliberately balanced set of variants around the same underlying buying question and run all of them, not just the original. If your brand appears across most variants but not the screenshot's phrasing, you have a phrasing artifact, not a visibility problem.
The failure modes here are systematic enough to have their own catalogue: see prompt-set mistakes for the seven that most often produce a score nobody should trust, including the ones that flatter you rather than hurt you.
A warning that cuts both ways. The temptation at this point is to write prompts that produce a nicer answer and report those instead. That is not diagnosis, it is score manufacturing, and it will fail the first time someone senior types their own phrasing into a chat window. Fix the set to be representative, which sometimes means keeping prompts that show you badly.
Cause 3: The Model Or Its Retrieval Changed
Third in likelihood, and the one people reach for first because it externalises the problem.
Providers update models, change default versions, adjust retrieval and grounding behavior, and modify product rules. Any of these can shift which brands get named, without anything changing on your side. It is a genuine cause. It is just far less common as an explanation for a specific single screenshot than the two above.
The test. This one is only runnable if you have a baseline, which is the whole argument for building one before you need it.
- Check whether the shift is category-wide or specific to you. Run the same prompt set and record every brand named, not just yours. If several established names dropped out simultaneously and a different set of names appeared, that pattern points at the system rather than at you.
- Check whether the model version or default changed between your baseline runs and now. Where the API exposes version identifiers, this is a direct comparison. Where it does not, note the gap honestly.
- Check whether retrieval behavior changed — for instance, whether source links are now returned on prompts that previously produced none.
- Check dated provider documentation and changelogs for the relevant window.
How to read the result. A category-wide simultaneous shift with a documented provider change behind it is a strong case. A category-wide shift with no documentation is worth writing down and watching, stated as an observation with an unknown cause rather than as a conclusion.
What not to do. Do not assert a model change to leadership because it is the most comfortable explanation available. If you cannot show the category-wide pattern, you do not have the finding.
Cause 4: A Genuine Visibility Gap In The Sources It Uses
Least common as the explanation for one screenshot, most consequential when it is real.
This is the case where, across many runs and many representative phrasings, over a stable period, your brand is genuinely absent or thinly represented while competitors are consistently named. Underneath it usually sits something concrete: your company is described inconsistently across the public record, or it is absent from the kinds of third-party sources that get retrieved for your category, or its own material does not clearly state what it does, for whom.
The test. You only reach this test after Causes 1 to 3 have been ruled out with data.
- Establish that absence is consistent across runs, across phrasings, and across at least two surfaces.
- Capture what sources are being cited when the competitor is named, where citations are available. That set is the observable part of the picture and the most useful artifact this whole exercise produces.
- Look at how your company is described in those source types, if at all, and whether that description matches how you describe yourself.
- Compare against a competitor systematically rather than impressionistically — the protocol in the competitor teardown is designed for exactly this step.
A caution before you act on it. Nobody outside the providers has a validated causal model linking a specific publishing action to a specific change in answer composition. You will find work that is plausibly worth doing on its own merits. Treat it as that, on a stated hypothesis, with a measurement plan attached, and not as a lever that reliably reverses a competitor's position. Anyone selling you certainty here is selling you something they do not have.
How To Tell Which One It Is, In Order
Run these in sequence. Stop when a cause is confirmed. Do not skip ahead because a later cause is more interesting.
- Re-run the exact prompt many times. If your brand appears at a reasonable rate, stop. It was sampling. Report that.
- Test balanced phrasing variants. If your brand appears across most variants, stop. It was the phrasing. Fix the set and report that.
- Check for a category-wide shift and a documented provider change. If the whole category moved together, you have a system-side explanation. Record it and keep watching.
- Only now, investigate the source picture. Consistent absence across runs, phrasings, and surfaces is the threshold for calling it a genuine gap.
Two rules govern the whole sequence.
Baseline before spend. No remediation budget should be approved before you have a documented baseline: a frozen prompt set, a recorded sample size, named surfaces and model versions, run dates, and the variability you observed. Without it you cannot tell later whether anything you did worked, which means you will be having this same conversation next quarter with the same absence of evidence. The full standard is on our methodology.
When the honest answer is "this is noise, wait," say it. If your re-runs show your brand appearing at a rate consistent with what you saw before, the correct output of this exercise is a short note saying the observation is within normal variation, here is the sample size, here is the rate, we are continuing to monitor. That is a complete and professional answer. Saying it is more useful than inventing a project, and telling a real change from noise is a solvable problem with a method behind it. See reading variance.
What To Do About Each, And What Not To Do Yet
| Confirmed cause | Do this | Do not do this yet |
|---|---|---|
| Sampled once | Report the observed rate with its sample size. Set a monitoring cadence. | Launch a project. Change the site. Escalate. |
| Prompt phrasing | Rebuild the prompt set to be representative, version it, freeze it, re-baseline. | Swap in prompts that produce nicer answers. |
| Model or retrieval change | Document what moved and when, keep measuring, wait for the pattern to stabilise. | Attribute a competitor's gain to their marketing. Rebuild strategy on one shifted run. |
| Genuine gap | Write down the specific hypothesis, the work it implies, and how you would measure whether it did anything. | Promise leadership a timeline for reversal. Buy a tool before you know what you are tracking. |
One more thing not to do, in all four cases: nothing on this site will tell you how to manipulate an assistant's output, and the tactics circulating for doing so — coordinated fake reviews, astroturfed community posts, content aimed at injecting instructions into retrieval — are liabilities rather than levers. They are covered here only as risks to detect.
Where To Go Next
The supporting posts in this cluster each go deep on one step of the sequence above.
- Why Two People Asking The Same Question Get Different Brands — the mechanism behind Cause 1, and why a single answer is never a measurement.
- Prompt-Set Mistakes That Make Your Visibility Score Meaningless — seven ways a prompt set quietly flatters you, for Cause 2.
- Reading Variance: Telling A Real Change From Noise — how to decide when a movement is worth acting on.
- Nine Ways AI-Visibility Reports Mislead Executives — how to audit a report before you present it, including your own.
- How To Run A Competitor AI-Visibility Teardown In An Afternoon — a repeatable protocol for reading your category once Cause 4 is on the table.
In Plain English
One screenshot of a competitor being recommended is a single sample from a system that gives different answers to the same question. Most of the time it means nothing. Before you conclude anything, re-run the prompt many times, then test fairer phrasings, then check whether the whole category moved, and only then investigate whether you are genuinely thin in the sources the system draws on.
This is for whoever has to answer the escalation, and for anyone about to approve spending against a problem that has not been demonstrated. Use it the day the screenshot arrives, and again as the standing procedure whenever a number moves. Do not use it as a promise that any of this reverses a competitor's position, because no method can promise that.
Your next step is a baseline: one representative prompt set, run properly, with the sample size and dates written down.
Run the free AI Visibility Check to get a sampled, multi-model snapshot with an honest confidence range attached to it, rather than a screenshot.
Primary-source verification
Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.
- Google Search: AI features and your website - How Google describes eligibility, preview controls, and appearance in its AI search features.
- Google: common crawlers and special-case user agents - Current Googlebot and Google-Extended identities and controls.
- OpenAI: overview of OpenAI crawlers - Current GPTBot, OAI-SearchBot, and ChatGPT-User descriptions and publisher controls.
- Anthropic: web-crawler controls - Current ClaudeBot, Claude-SearchBot, and Claude-User descriptions and controls.
- Perplexity: crawler and user-agent documentation - Current PerplexityBot and Perplexity-User descriptions and controls.
- FTC: Endorsement Guides Q&A - Current U.S. regulator guidance on endorsements, disclosures, and reviews.
- FTC: Consumer Reviews and Testimonials Rule Q&A - Current U.S. rule guidance on deceptive review practices.