Why ChatGPT, Gemini and Perplexity pick different winners
Same question, three different shortlists. The reasons are structural rather than random, and they tell you which engine is worth fighting for first.
The first thing most teams notice when they start measuring AI visibility across engines is that the scores don't agree. A brand can sit comfortably in the top three on one engine and be entirely absent from another, for the same question, on the same day. This is usually read as evidence that the measurement is broken. It isn't. It's evidence that these systems are built differently and are answering slightly different questions.
Understanding why the spread happens is what turns a confusing dashboard into a plan. There are four structural differences that account for most of it.
Difference one: how much they lean on live retrieval
Some engines answer largely from the model's internal knowledge and reach for the live web only when the question demands it. Others search first and generate second, treating the model as a summariser of freshly retrieved pages. This single design choice produces most of the variance you'll see.
A retrieval-heavy engine rewards recency and page-level quality. Publish a genuinely good page answering a specific question and you can appear within days. A knowledge-heavy engine rewards accumulated presence: how often you've been discussed, referenced and described across the web over a long period. You cannot shortcut that with a publishing sprint.
If your scores are high on the retrieval-heavy engines and low on the others, you have good content and a weak footprint. If it's the reverse, you're well known and your pages aren't doing their job. Those are opposite problems with opposite fixes.
A wide spread between your best and worst engine is more informative than the average. The average tells you how you're doing. The spread tells you which half of the work you've been neglecting.
Difference two: which corner of the web they trust
Every engine has an implicit hierarchy of sources, and they are not the same hierarchy. Some lean heavily on community discussion and forums. Some prefer established publishers and reference sites. Some weight the search index they inherited from a parent product. Some visibly favour recent, well-structured commercial content.
The practical consequence is that where you are talked about matters as much as what you say about yourself. A brand with an active, organic community presence will do disproportionately well on engines that trust community sources, even with a mediocre website. A brand with excellent documentation and no third-party footprint will do well on retrieval-heavy engines and vanish elsewhere.
This is worth checking rather than assuming. When we look at citation sources across engines for the same brand, the overlap is often smaller than people expect — the pages earning you a mention on one engine frequently aren't the pages doing it on another.
Difference three: how many names they're willing to give
Answer length is a product decision, and it changes the competitive maths entirely. An engine that habitually returns three options is a far harsher environment than one that returns eight. Being fourth-best in your category is invisible in the first and perfectly fine in the second.
This also affects how you should read a drop. Falling out of an eight-name list is a real decline. Falling out of a three-name list might just mean the engine trimmed its answer format that week. Without knowing the typical list length for each engine, you'll misinterpret half your movement.
Difference four: how the question is interpreted
Ask "what's the best tool for tracking AI visibility" and one engine will interpret it as a request for a ranked recommendation, another as a request for an overview of the category, and a third as a request for the criteria you should use to choose. Three different intents, three different sets of brands surfaced.
Prompt phrasing therefore does more work in AI visibility measurement than keyword selection ever did in search. Small changes in wording produce large changes in results, which is both a measurement hazard and an opportunity: it means the questions your buyers actually ask are worth tracking precisely, in their exact phrasing, rather than as a tidy category term.
So which engine should you optimise for?
The tempting answer is the biggest one, and it is usually the wrong answer. Three considerations matter more than raw user numbers.
Where your buyers actually are
Developer tools and consumer retail have very different engine mixes. Ask a handful of recent customers what they used. It's a five-minute exercise that beats any assumption.
Where you're closest to the cut
Going from fifth to third on an engine where you nearly appear is far cheaper than breaking into one where you're absent. Fix the near-misses first.
Which fixes transfer
Crawl access, clean structure and consistent entity data help everywhere. Do those before anything engine-specific, because they're the only work that compounds across all of them.
There is also a strong argument against chasing any single engine too hard. The landscape is unstable — retrieval behaviour, answer formats and source preferences all change without notice, and a strategy tuned tightly to one system's current quirks ages badly. The durable work is the boring work: be reachable, be clearly described, be corroborated elsewhere, be genuinely useful on the specific questions your buyers ask.
Track all the engines, because the spread is the diagnostic. Optimise for the fundamentals, because they're what survives the next format change.
- →Disagreement between engines is structural, not a measurement fault.
- →The spread between your best and worst engine diagnoses which half of the work you've skipped.
- →Fix near-misses before absences — it's the cheapest movement available.
- →Engine-specific tactics age badly. Reachability, structure and entity clarity compound.