A person at a wooden desk comparing two computer monitors placed side by side.

Photograph by Lisa Fotios via Pexels. Illustrative image, not a customer or endorsement.

Blog · · 7 min read · NEXUM Team

Why ChatGPT, Gemini, Perplexity and Claude Don’t Recommend the Same Brands

If your team only asks ChatGPT whether you show up, you do not have an AI visibility program. You have one chat window.

That would not matter if the major assistants agreed. They often do not. The same buyer question can produce a different shortlist in ChatGPT, Gemini, Perplexity and Claude. Category leaders may appear almost everywhere; mid-market and challenger brands can look healthy in one engine and almost absent in another.

This is not a reason to panic, and it is not a reason to buy a separate “optimization” package for every logo on a slide. It is a reason to stop averaging the problem into a single score, and to measure the engines your buyers actually use.

Why “AI search” is the wrong unit

These products do not share one index, one crawler or one habit of citing sources.

Google has published its own guidance for showing up in AI Overviews and AI Mode. That document is about Google’s features. It does not tell you how ChatGPT gathers sources, and it does not describe Perplexity’s live-web habit. Treating “GEO” as one algorithm is how teams end up celebrating a good ChatGPT screenshot that their customers will never see.

A public 2026 test by a practitioner, who ran thousands of API calls across ChatGPT, Claude, Gemini and Perplexity for a set of SaaS brands, made the same point in numbers: a large share of pairwise comparisons between engines showed statistically meaningful differences, and the spread was wider on switching-intent prompts than on beginner questions. That study is not a census of the web, and you should not treat any one cell in someone else’s table as your forecast. It is a clear illustration of a pattern many teams already notice in their own niches: visibility is fragmented, and an average hides the fragment that matters.

Perplexity, in particular, tends to lean on what it can retrieve from the live web. Google’s AI features lean on Google’s index and policies. Claude often names fewer tools in a single answer. ChatGPT, when it searches, has its own retrieval path — Ahrefs’ work on 1.4 million ChatGPT prompts is about that path, not about Gemini. If you only study one of these, you will overfit.

Over-the-shoulder view of a person working across a laptop and two monitors in a dimly lit room.

The same question can live on more than one surface.

Photograph by Mikhail Fesenko via Pexels. Illustrative image, not a customer or endorsement.

Mentions, citations and “presence” are still not the same

Before you compare engines, agree on the event you are counting.

A brand can be mentioned in the answer, cited as a URL, both, or neither. An assistant can recommend you while citing a review. It can cite your article and recommend a competitor. If you collapse that into “we appeared,” you will not know what to fix. We walk through the measurement model in AI share of voice, including why a presence rate without position, sentiment or accuracy is a weak KPI.

The engine split makes that discipline more important, not less. A competitor can “win” Gemini as a named recommendation and “win” ChatGPT as a cited source. Those are different jobs.

What actually differs, in practice

You do not need a research budget to see the shape of this. You need a fixed list of questions and a way to log the answers.

What usually changes between engines is not a secret on-page tag. It is:

  • Whether the engine searched at all. A memorized answer and a retrieved answer are different systems. If ChatGPT does not search, there are no citations to inspect.
  • Which sources it already trusts. Wikipedia, YouTube, documentation, retail, Reddit and trade press do not have the same weight everywhere, and that mix moves over time.
  • How many brands it is willing to name. A conservative answer with three names and a generous one with ten will produce different presence rates even when your relative position is stable.
  • How fresh the retrieval is. A live-web engine can surface a recent comparison post; a more conservative one may keep recommending the incumbent it already knows.
  • The wording of the prompt. “Best [category] tool” and “I want to switch away from [incumbent]” are not the same intent. Fragmentation is often worse at the bottom of the funnel, which is exactly where a shortlist matters.

None of that is an argument for writing four versions of every page. A clear, factual page that answers a real question is still the right on-site unit of work. It is an argument for testing the same page against several answers, and for earning mentions in the sources each engine actually cites.

How to measure the gap without fooling yourself

Use one prompt set for every engine you care about. Do not invent a second list that is kinder to the model you already like.

A workable baseline:

  1. Choose the engines your buyers use. For many B2B and ecommerce teams in 2026, that starts with ChatGPT, Gemini and Perplexity. Claude matters in some professional workflows. If a surface is not in your product or your process yet, do not report it as if it were. (NEXUM currently measures ChatGPT, Gemini, Perplexity and Claude. We do not treat Microsoft Copilot as a measured surface in the product today.)
  2. Write prompts from sales and support, not from a keyword tool. “Best X for Y,” “X vs Z,” “alternative to [name],” and “is [your brand] right for [job]” will tell you more than a generic head term.
  3. Sample more than once. If you run a prompt three times and only count the best answer, you are not measuring visibility. You are editing a highlight reel. Citation instability is documented.
  4. Log mention, citation, who else is named, and whether the description is accurate. Wrong pricing or an outdated feature can be worse than absence.
  5. Report each engine separately, then a range or a simple average if you must. If you are at 50% in ChatGPT and 10% in Perplexity, the average is a political number, not a diagnostic one.
  6. Keep referral traffic in a different column. A click from chatgpt.com or perplexity.ai is useful and incomplete. Many people who take a recommendation later arrive as branded search or direct traffic. That is a measurement problem, not proof that the mention had no effect.

If you want a longer playbook for the on-site work that supports retrieval, use how to get cited by ChatGPT, Perplexity and Gemini. The point of this article is narrower: do that work with a panel that can show you when one engine disagrees with the others.

What to do when the shortlists disagree

Start with the engine that your customers actually open, not the one where the screenshot looks best.

If you are absent in a live-web engine but present in a more “memory-heavy” one, look at recent third-party pages and at whether those URLs are crawlable. If the opposite is true, the issue may be older entity reputation, not last week’s blog post.

If descriptions are wrong in one assistant and fine in another, find the source that assistant is citing and fix the factual error there — your pricing page cannot override a stale profile if the model is not looking at your pricing page.

If you only lose on switching-intent prompts, you probably do not have a “visibility” problem in the abstract. You have a comparison problem: the model can describe you, but it will not put you on a shortlist against a named incumbent. That is usually a content and reputation job (honest alternative pages, independent reviews), not a robots.txt job.

And if the answers move every week, shrink the prompt set until you can re-run it. A dashboard with 400 prompts you never repeat is worse than 25 you measure properly.

Three colleagues at a table, one pointing at a whiteboard diagram while the others listen and take notes.

When shortlists disagree, the work is to compare them — not to average them.

Photograph by PNW Production via Pexels. Illustrative image, not a customer or endorsement.

A note on tools and guarantees

You can do this in a spreadsheet. You should, at least once, so you know what the work looks like. A product is useful when you want the same panel run on a schedule, across several assistants, with mentions and citations stored so you can see change. It is not useful if it outputs one “AI score” and a promise that the next answer will include your name.

AI mentions cannot be guaranteed. The useful promise is narrower: measure the same questions, in the engines that matter, and work the gaps you can actually see.

If you want help with that baseline, start with a free NEXUM audit. We will show you how those assistants currently talk about your brand and a first list of fixes. We will report the engines separately.

FAQ

If I am visible in ChatGPT, will I be visible in Gemini?

Not necessarily. Do not assume spillover. Test the same prompts in each engine you care about.

Should I write different pages for each assistant?

Usually no. Write clear pages for buyers. Then see which engines retrieve or name you, and earn the off-site sources each one cites.

Why does NEXUM not include Copilot in this list?

Because we should only claim what we measure. The product today covers ChatGPT, Gemini, Perplexity and Claude. Copilot is a relevant surface for some buyers; it is not one we will report here as if it were already in the product.

Is a single “AI share of voice” number wrong?

It is incomplete. Use it, if you must, alongside the per-engine numbers and the mention/citation split. The average should not be the only slide.