Research · AI recommendations
How often does AI actually recommend a brand?
We read 152,195 brand mentions in 65,746 AI answers and sorted each one by what the answer does with the brand. About one in a hundred is a top pick. Most brands are simply listed.
October 202665,746 answers1,509 prompts5 AI engines
Definitions
What counts as a recommendation?
Most visibility metrics count every brand in an AI answer the same way. But an answer can do very different things with a brand name. It can pick the brand as the one to choose, praise it for a certain need, list it as one option among several, mention it in passing or only show it in a source link.
We sorted every brand mention into one of these five types. In this article, a recommendation means a top pick or an endorsement. The examples below are made up to show what each type looks like.
If you just want one, go with Brand A.
The answer singles out the brand as the one to choose for the question that was asked.
Brand B – best for small teams.
Called the best for a specific need, or named among the best, but not singled out as the one to choose.
Brand C – fixed-price packages and weekend viewings.
One of several options in a list, a table or a sentence, without praise that sets it apart.
…integrates with accounting tools such as Brand D.
Named as background – an example, an integration, a parent company, a comparison or a warning.
[Brand E – Pricing and plans]
The brand only appears in the text of a source link, not in the answer itself.
481 brand mentions reviewed one by one, from all five engines and all types of prompts, 15 Sep – 11 Oct 2026. Top picks are 1.5% of the mentions, too small to label in the bar.
Only 1.5% of brand mentions are top picks, and 16% are endorsements. Most brands, 59%, are listed as one option among several. That is from 481 mentions we reviewed one by one. When we classified all 152,195 mentions automatically, the shares came out at 1.0% top picks and 15.5% endorsements.
Brands in answers
How often do brands show up in AI answers at all?
In about half of them. 53% of all answers name at least one brand, four on average. When the prompt asks for a recommendation, 68% of answers name a brand. Even general questions bring up brands: 39% of answers to questions like how a fund fee works or what to think about when buying a home name at least one.
The engines differ here too. On the prompts all five answered, ChatGPT names a brand in 74% of its answers and Gemini in 69%, while Perplexity does so in 58% and AI Mode in 56%. We looked at how many brands each engine names per answer in an earlier study: How many brands does AI mention in its answers?
Recommendations
How often does an AI answer actually recommend a brand?
Rarely with a clear pick. Only about one in 20 answers that name a brand makes a top pick. Our automated count finds a top pick in 4% of them, and our checks suggest the true share is a little higher, around 5%. Another 27% endorse one or more brands without picking one. The rest recommend no brand at all: they list the options and leave the choice to the reader. Our automated count puts that at 69%, but when we checked 40 of those answers one by one, 4 did recommend a brand, so the real share is closer to six in ten.
Asking for a recommendation helps, but not much. When the prompt asks for one, 7% of the answers that name a brand make a top pick, 35% only endorse and 59% recommend nothing. Figure 2 shows how much this varies by engine. For general questions, a top pick is very rare (1%), but 15% of the answers still endorse a brand. One in six of all recommendations in our data comes from a prompt that didn’t ask for one.
Answers that name at least one brand, for the 191 prompts that ask for a recommendation and that all five engines answered. Automated count. Our checks suggest the rules miss some recommendations, so the grey “no clear recommendation” share is about 5 points too high. The differences between engines hold.
Engines
Do the five engines recommend the same way?
No. To compare the engines fairly, we used the 339 prompts that all five answered. ChatGPT is the most willing to pick. It makes a top pick in 10% of the answers where it names a brand, and recommends at least one brand in almost half of them. AI Overview and Perplexity make a top pick in 6–7% of those answers.
Google AI Mode almost never picks. It makes a top pick in under 1% of its answers that name a brand, 80% of them recommend no brand at all, and only 9% of the brands it names are endorsed or picked. Gemini endorses brands about as often as Perplexity, but rarely picks one (2%).
ChatGPT also differs in another way. 9% of the brands in its answers only appear as source links, because it often writes the title of a source page into its answer. On the other engines, that share is 1–5%.
| Engine | Answers naming a brand | Brands per answer | Answers with a top pick | Answers that only endorse | Mentions endorsed or picked | Source only |
|---|---|---|---|---|---|---|
| ChatGPT | 74% | 4.4 | 10.0% | 38% | 24% | 9% |
| Perplexity | 58% | 3.9 | 6.1% | 32% | 24% | 3% |
| Google AI Overview | 61% | 3.8 | 6.5% | 22% | 16% | 1% |
| Google AI Mode | 56% | 4.2 | 0.6% | 19% | 9% | 5% |
| Gemini | 69% | 4.5 | 1.9% | 32% | 16% | 3% |
| All five engines | 65% | 4.2 | 5.6% | 30% | 19% | 5% |
The 339 prompts all five engines answered (18,151 answers). Answer shares cover answers that name at least one brand. Automated count.
Agreement
Do the engines agree on which brand to recommend?
Mostly not. When two engines answer the same prompt on the same day, they name the same brand first only 28% of the time. When both recommend something, they recommend at least one of the same brands 59% of the time. ChatGPT and Perplexity agree most often (64%), and AI Mode and Gemini least often (46%, based on 128 cases).
Top picks agree more often, but mostly for a few prompts. In the 301 cases where two engines both made a top pick for the same prompt on the same day, they picked the same brand 65% of the time. Those cases come from only 62 prompts, and five of them account for 43% of the cases – prompts where one brand clearly leads the category, such as the largest platform in a market. Counting each prompt once, the engines pick the same brand about half the time.
It is rare for all engines to recommend anything at all. We looked at 9,545 cases where ChatGPT, Perplexity, AI Overview and AI Mode all answered the same prompt on the same day. In 70% of them, none of the four recommended a brand. All four recommended something in only 101 cases (1%), and in 42 of those they had a brand in common. All four made a top pick in just 2 cases.
Same prompt, same day, where both engines recommend at least one brand (top pick or endorsement). “First” shows how often both engines name the same brand first, whether or not they recommend it.
Position
Does position in the answer affect recommendations?
Yes. 91% of top picks are the first brand named in the answer. That is partly how answers are written: an answer that picks one brand usually leads with it. It holds even when we only count picks made with explicit phrases like “my top pick” or “I’d go with”, wherever they appear in the answer: 86% of those are the first brand named.
Endorsements follow the same order, less steeply. Across all answers, the brand named first is endorsed or picked 25% of the time, and the brand named tenth 8% of the time. When the prompt asks for a recommendation, it drops from 34% to 9%. For general questions, position barely matters (7–10%), because those answers rarely rank brands at all.
Position is the order in which brands first appear in the answer. 152,195 brand mentions.
Two of our classification rules look at the first sentences of the answer, which could favour the first brand named. With those rules switched off, the pattern stays the same: 21% of first-named brands are endorsed or picked, dropping to 8% at position ten.
Citations
Are brands that get cited more likely to be recommended?
Yes, and it matters which site is cited. Brands that are cited are endorsed or picked more often than brands that aren’t. Part of that is because cited brands tend to be named earlier in the answer, so we compared brands in the same position:
- The brand’s own site is cited. Slightly more recommendations. Named first, these brands are endorsed or picked 28% of the time, compared with 26% for brands that aren’t cited.
- The brand is named on another site’s cited page, such as a review, a comparison or a buying guide. Clearly more recommendations: 37% when named first.
- Both. The highest rate: 42% when named first, and about twice the rate of uncited brands further down the answer.
The gap is bigger for top picks. Named first and not cited, a brand is the top pick 4% of the time. With its own site cited it is 6%, named on another site’s cited page 8%, and with both 13% – about three times as often as an uncited brand.
120,131 brand mentions in answers that cite at least one source. “Named on another site’s cited page” means the brand’s name is in the title of a cited page on a website that isn’t the brand’s own.
Being named on another site’s cited page goes with more recommendations on all five engines: ChatGPT 25% compared with 14% for brands that aren’t cited, Perplexity 31% compared with 21%, AI Overview 25% compared with 16%, AI Mode 17% compared with 11% and Gemini 26% compared with 16%. The brand’s own site varies more by engine. It goes with many more recommendations on ChatGPT (24% compared with 14%), a few more on AI Overview, AI Mode and Gemini, and fewer on Perplexity (16.5% compared with 21%).
This shows a link, not proof that citations cause recommendations. Engines may cite comparison and review pages because those pages rank brands, and the answer then repeats that ranking. We also only look at page titles, so other sites that mention a brand further down the page aren’t counted. Still, the pattern is clear: a brand’s own site helps an engine describe it, while other sites are more closely tied to it being recommended.
Prompts
Which prompts lead to a recommendation?
The wording of the prompt matters most. When the prompt asks for “the best”, 50% of the answers that name a brand recommend one, and 8% make a top pick. Prompts that ask “which”, “where” or “who” get a recommendation in 31% of answers and a top pick in 5%. Prompts that ask “how”, “what” or “why” almost never get a top pick, and 12% get an endorsement.
An answer that starts with “it depends” still often recommends a brand. 13% of answers start that way, and they endorse brands more often than other answers (43% compared with 30%), usually one brand for each need.
34,879 answers that name at least one brand. Automated count. The one-by-one review shows the same pattern: 25% of brand mentions in answers to “best” prompts are endorsed or picked, compared with 14% for other prompts.
What it means
What should brands take from this?
- Being listed is the norm, a top pick is rare. About six in ten answers that name a brand recommend none. Count how often you are endorsed or picked, not only how often you are named.
- Give AI a clear reason to choose you. Most recommendations are “best for” a specific need. Make it easy for engines and the sites they cite to say what you are best for.
- Get named on other sites’ pages. Comparisons, reviews and buying guides that engines cite go with more endorsements and top picks than your own site being cited, and the two together go with the most.
- Check every engine. ChatGPT makes a top pick in one in ten answers that name a brand, AI Mode in fewer than one in a hundred. A recommendation on one engine says little about the next.
See which brands AI names in your category – and where you stand
Try VerciteMethodology
How we did it
- Data
- 65,746 answers from ChatGPT, Perplexity, Google AI Overview, Google AI Mode and Gemini to 1,509 prompts tracked in Vercite, collected 15 September – 11 October 2026. 34,879 answers name at least one brand.
- Brand mentions
- 152,195 mentions of brands, products and companies. If a brand is named in the prompt itself (for example “Is Brand A better than Brand B?”), its mentions in the answer are left out.
- Prompt types
- Prompts that use words like best, recommend, good, which, where or who count as asking for a recommendation (722 prompts). Prompts that name a specific brand count as questions about a brand (92). The rest are general questions, such as how something works or what it costs (695).
- Prompt mix
- About half of the prompts are in Swedish, most of the rest in English, with smaller sets in Norwegian, Finnish, Portuguese and Romanian. They come from brands tracked in Vercite and lean towards finance, real estate, hospitality, energy and business services. The results describe this set of prompts, not all AI use.
- Classification
- Each mention was sorted automatically by a set of rules that read the answer around the brand: lists and tables, headings, phrases such as “best for”, “my top pick”, “bäst för” and “förstahandsval”, and the first sentences. A top pick needs an explicit pick phrase in a sentence that names only that brand, or an opening that singles out one brand. We checked the rules against 481 mentions reviewed one by one across all prompt types, made by an AI reviewer reading the full answer, separately from the rules. The rules agree with the reviews 92% of the time on whether a brand is endorsed or picked. 80% of the mentions they mark as endorsed or picked are, and they find 77% of them. On 86 mentions set aside before the rules were tuned, agreement was 91%.
- Checking top picks
- Top picks are too rare for the 481 reviews alone: there were 7, and the rules found 3 of them. So we ran two more checks, also reviewed by AI. Of 40 randomly chosen top picks from the rules, 33 (83%) were real top picks. Of 40 random answers where the rules found no top pick, one had a top pick that was missed. Of 40 random answers where the rules found no recommendation at all, 4 (10%) did recommend a brand. The rules never marked a non-top-pick as a top pick among the 481 reviews. Together this suggests the rules are cautious, and the true share of answers with a top pick is around 5% rather than 4%. A separate review of 45 whole answers found a recommendation in 47% of them, compared with 38% for the rules, so the automated share of answers without a recommendation is on the high side.
- Weighting
- Shares are calculated over all answers and mentions together, so prompts that produce many brand mentions count more. If every prompt counts equally, the share of top picks drops from 1.0% to 0.8% and endorsements from 15.5% to 10.7%, but the patterns by position, prompt type and citation stay the same.
- Brand detection
- Brands are those found in each answer by Vercite’s brand detection. Brands it misses, or names that aren’t really brands, affect these numbers.
- Position
- The order in which each brand first appears in the answer.
- Engine agreement
- One answer per prompt, engine and day. Different spellings of the same brand name were matched.
- Citations
- “Own site cited” means one of the brand’s own websites is among the sources the answer cites. “Named on another site’s cited page” means the brand’s name is in the title of a cited page from another website. For ChatGPT, we only count sources it shows as citations, not every page it read.
