Three ways a brand "disappears"
When somebody says "we used to come up in ChatGPT and now we don't", we look for one of three things, and they need different responses. The three are our framing; the tracker measures the first and does not separate it from the second.
- The answer moved on its own. The same engine, asked the same question, names a different set of companies while nothing about the brand changed. The tracker figures below show how much names change between captures. They do not separate random variation from model or collection effects.
- The engine changed. A new model, a changed answer length, a different search route or a reworked interface can change which names come through. The brand did nothing; the instrument did.
- The brand's evidence changed. Fewer recent reviews, a competitor named in more comparison pieces, inconsistent company details across directories, or a site the engine's retrieval no longer finds. This is the one a marketing team can act on most directly, and it is worth checking the first two before assuming it.
How much AI answers move on their own
Our travel edition puts 20 fixed, unbranded travel questions to six engine services on a weekly tracker in a UK-configured market and matches the answers against a frozen list of 242 travel names. Coverage differs by engine: ChatGPT and Perplexity have only two or three captures in the window. The first edition covers August to September 2026, with 742 usable answers. It is a question panel, not a league table, and a name match is not a recommendation. Three findings bear on the "disappeared" question.
Engines disagree with each other. In the 28 September six-engine snapshot, 52% of the names matched per question appeared in just one engine's answer, on average (53% in the 21 September capture). For one question, "What's the best tour operator for a first safari in Africa?", the six engines named 20 different names between them on 28 September. Ten appeared in only one engine's answer. None appeared in all six. Seeing a name in one engine's answer cannot establish its visibility across the others; a single engine gives a partial view.
The same engine changes its mind between captures. For the repeat comparisons, the tracker asks the same engine the same question in the next capture and divides the names in both answers by the names in either. Mean overlap was 48% for Google AI Overviews (117 pairs), 47% for Google AI Mode (139 pairs) and 55% for Gemini (80 pairs). The two answers shared only about half of their combined names. Pairs affected by known model, answer-limit or collection changes were excluded, as was a pair collected less than a day apart, and Google capture settings are not fully recorded. The edition is careful to say the data shows answers changed, not why, and that different periods and answer lengths apply, so the three figures are not an engine quality ranking.
Being mentioned first is not a held position. The name an engine mentioned first usually came back in the next capture, 83% to 91% of the time depending on the engine. It was first again far less often: 50% to 60%. The two measures have different denominators, and "first" means the earliest matched name in the captured text, not the engine's preferred choice.
The repeat figures cover Google AI Overviews, AI Mode and Gemini only. ChatGPT and Perplexity had too little comparable history in the window, and Claude's collection route changed, so none of the three is in the repeat charts. We would not assume ChatGPT is steadier than Google and Gemini; we have not measured it yet.
Why ChatGPT in particular
ChatGPT reaches a brand two ways: what it saw while it was trained, and what it finds when it searches for a live question. ChatGPT's live search uses Bing, so what Bing returns can shape the answer, and a brand that was not discussed in the comparisons, reviews and forum threads the model trained on has little to be recalled from. We go through the mechanism in how ChatGPT decides who to recommend.
For the "disappeared" case, the point is narrower. Both routes draw on pages the brand does not control, and both can change without the brand doing anything: Bing re-ranks, a competitor earns a new round-up, the model is updated. These are possible mechanisms, not something the tracker measures. Until we have repeat-capture figures for ChatGPT, the honest reading of a missing ChatGPT mention is the same as for Google and Gemini: check whether it is still missing next week, and in the other engines, before deciding it is a loss.
How to tell variation from loss
- Keep the question fixed. Changing the wording can change the answer. Compare like with like, and record the market and the collection date.
- Ask more than once. One answer per engine per capture, on a fixed schedule. In our judgement, a brand missing from one capture is weak evidence of a loss until it recurs.
- Ask more than one engine. With half the names appearing in a single engine's answer, absence from one engine says little about the others.
- Read the context. A name match can be a recommendation, a passing mention, a warning or a link. Count the kind of mention, not just the mention.
- Log what changed. Model versions, collection settings and your own site changes, with dates. A fall that lines up with a model update may not be a fall your content caused. A higher number alone does not show your intervention worked either.
If a brand stays missing across captures and engines, with no instrument change we can identify, the third cause becomes more plausible. Our tracker cannot confirm it. Our post on why brands are missing from ChatGPT answers goes through seven causes and a fix for each.
Frequently asked questions
Why did my brand disappear from ChatGPT recommendations?
Possibly it did not. In our tracker, consecutive answers to the same question shared only about half of their combined names on the engines we could compare (Google AI Overviews, AI Mode and Gemini), and ChatGPT is not in those comparisons. Check the same question over several weeks and in other engines first. If the brand stays missing everywhere, look at the third-party evidence ChatGPT retrieves and trains on, starting with what Bing returns for your category and the comparison pages it finds.
How often do AI answers change?
From one capture to the next, in our data. Mean name overlap between consecutive answers to the same travel question was 47% to 55% depending on the engine (Google AI Overviews, AI Mode and Gemini, August to September 2026, 336 eligible answer pairs). The edition reports that answers changed, not why.
If we are mentioned first, is that a held position?
Not in our data. The first-mentioned name came back in the next capture 83% to 91% of the time, but was first again only 50% to 60% of the time (different denominators). A first mention is also not a recommendation by default; it can be a comparison, a warning or a link.
Sources
- SearchIntel, The Answer Economy Tracker, Travel 01: 742 usable answers to 20 fixed travel questions across six engine services, UK market, August to September 2026. Data to 28 September 2026, prepared 10 October 2026. Method, exclusions and downloadable tables on the page.
- SearchIntel, How ChatGPT decides who to recommend.