Short answer: nobody outside an AI engine can see how it chooses which brands to name, and any guide that claims to know the exact mechanism is guessing. What you can observe is the pattern: whether answers in your category name anyone at all, how concentrated the names are, how much the engines differ, and which sources they cite. Those patterns, plus a few preconditions you fully control, are enough to decide where to work.
This guide separates what can be measured from what is often asserted. The category figures come from our own Q3 2026 benchmark of 12 categories, five models and 1,188 answers, published as the AI recommendation benchmark.
What can and cannot be known
An assistant composes each answer from what it learned in training and, when retrieval is on, from pages it fetched for that question. You can record the output: which brands were named, in what form, and which sources were marked. You cannot record the reasoning. If a competitor is named and you are not, the observable fact is exactly that. Why the engine chose it is not something answer data supports, and inventing a reason tends to send the work in the wrong direction.
That is why the useful questions are about patterns rather than about the engine's mind. Each of the patterns below is something you can check for your own category in an afternoon.
Five patterns in how AI assistants name brands
Each pattern below can be checked in your own category with a fixed set of questions and a few engines. None of them explains the engine’s reasoning; together they tell you where effort is likely to be wasted and where it is not.
Pattern 1: not every answer names a brand
Before asking how to beat a competitor in an answer, check whether the answer names anyone. For some buyer questions the assistant names no provider at all. Those questions are a different kind of opportunity from the ones where a competitor appears and you do not, and they should be tracked separately: on open ground nobody has to be displaced, so becoming the first concrete answer is cheaper than replacing an incumbent.
Two lists, not one: questions where a competitor was named and you were not, and questions where nobody was named. Mixing them produces a to-do list that cannot be prioritised.
Pattern 2: concentration depends on the category
In our benchmark, three categories had what we call a default answer, meaning one provider appeared in a quarter or more of all answers: project management software at 38% of answers, skincare at 35% and electric bikes at 28%. All three are product categories. Every local-service category we measured looked different, with the most-named provider below 13%.
The two situations call for different work. In a consolidated category, merely being present is not enough; the realistic target is a specific sub-question the incumbent answers badly, such as a budget, a constraint or a use case. In a fragmented category, a well-documented business is competing against a long tail rather than against an entrenched name. Per-category results are on our industries pages.
Pattern 3: the question changes the list
Assistants answer situations. A question that carries a budget, a constraint or a use case produces a different set of named brands from the bare category term. That has two consequences. First, you can be absent for "best project tool" and present for "project tool for a small agency that bills by the hour", and both facts matter. Second, any measurement built on keyword fragments rather than full questions describes a situation buyers rarely put to an assistant. The intent behind a question matters too: being named in research questions and absent from purchase questions is a different problem from the reverse. How to write questions that reflect this is covered in building a prompt set.
Pattern 4: the engines differ enough to matter
On the same questions in our benchmark, one model named 3.5 providers per answer and another named 2.1. A brand can be comfortably present in the engine that lists many options and invisible in the one that lists few, and a single cross-engine average describes neither situation.
Even two surfaces from the same company diverge. Across 540,000 query pairs, Ahrefs found that Google’s AI Mode and AI Overviews cited the same URLs only 13.7% of the time, although their answers reached 86% average semantic similarity. The analysis compared single generations of each, so part of that gap is ordinary variation between runs; either way, the sources behind similar-sounding answers were mostly different.
Citation behaviour differs as well. Perplexity exposes sources on nearly every answer, so a block on its crawler shows up in citation data immediately; its bot documentation separates the indexing agent, PerplexityBot, from the user-triggered Perplexity-User. Google draws a different line again: Google-Extended is not a crawler but a usage control, and blocking it removes your content from grounded Gemini answers while, according to Google's crawler documentation, leaving Google Search inclusion unaffected. Read results engine by engine before you read any average.
Pattern 5: the cited sources are often not the brands
In the product categories we measured, a large share of citations went to marketplaces, review sites and publisher round-ups rather than to brand sites. A brand's product page can be accurate and well built and still lose the citation to a listing page that carries the same facts with third-party framing. Comparison and round-up pages match the shape of a buying question directly, which is why they are prominent in citation lists for commercial categories.
Which source leads also depends on the engine. In Profound’s analysis of 680 million citations from August 2024 to June 2025, Wikipedia was the most-cited domain in ChatGPT at 7.8% of all citations, while Reddit led in Perplexity at 6.6% and in Google AI Overviews at 2.2%. The shares are small, which is itself informative: citations are spread across many domains, so the list worth acting on is the one cited for your own buyer questions, not a platform someone says is important everywhere.
The related point is about confidence. A model needs to see the same fact in independent places before it will state it confidently, so a claim that exists only on your own domain is the weakest evidence you can offer, however well it is marked up. Third-party corroboration covers how to find the sources that matter in your category and check what they say about you.
The preconditions you do control
None of the patterns above can be forced, but three preconditions decide whether you are eligible at all:
- Access. The search and user-triggered agents must be able to fetch your pages: OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot and Perplexity-User for Perplexity, Claude-SearchBot and Claude-User for Claude. xAI publishes no documented crawler user agent, so for Grok the only thing you control is that pages are readable to general crawling.
- An unambiguous entity. Spelling variants, a product line described as if it were a separate company, or a legal name that differs from the trading name can each split you into several rows. Organization markup with sameAs links to your real profiles is one of the few identity signals you can publish without waiting for anyone else; see entity and sameAs.
- Facts that can be lifted. Assistants lift sentences that answer a constraint: a dimension, a material, a return window, a compatibility. Copy written for brand tone gives them nothing to lift, however well it reads.
How to read which brands AI names in your own category
| Check | What it tells you | What to do with it |
|---|---|---|
| Questions where nobody is named | Open ground | Answer those questions concretely on pages agents can fetch |
| Share held by the most-named provider | Whether there is an incumbent | If there is, aim at a sub-question it answers badly |
| Names per answer, per engine | Where a list has room for you | Report and prioritise engine by engine |
| Which domains are cited | Who the engines currently trust for your category | Make sure those sources carry your current facts |
| Questions where a competitor is named and you are not | Your concrete gaps | Read the original answer before writing anything |
Run the same questions on the same engines again before drawing conclusions. Answers vary between runs even when nothing has changed, so a single run shows a snapshot, not a position.
Sources
- Our own Q3 2026 AI recommendation benchmark (Figures 1 and 2).
- Ahrefs, AI Overviews compared with AI Mode, December 2025.
- Profound, citation patterns across AI platforms.
- Perplexity, bot documentation; Google, common crawlers; OpenAI, crawler overview.
An earlier version of this guide quoted further aggregate figures from the same benchmark that are not in the fact files this page is checked against, so they were removed rather than restated.