Rankfor.AI research · August 2026

You might assume AI assistants read the same sources.
They do not.

119,055 citations from 78 production scans of 44 brands, in 16 languages, 13 July 2026 to 12 August 2026.

Across all Rankfor studies: 167,551 citations analysed.

GeminiChatGPTGrok

Live customer scans. Brands are anonymised and the corpus is aggregated, so no company is named here.

Where do answers come from?

Most of what AI cites about a brand is somebody's company website.

Sort every citation into 6 groups. Company pages, the brand's own and other companies', take 44.71%. Editorial media, national and regional news, trade and consumer press, blogs, and wire releases together take 20.95%.

National news, the category a communications budget is usually judged on, is 0.77% of citations.

A few long scans pull that 20.95% average up. The median brand gets 11.31% of its citations from the press. The middle half runs from 8.56% to 21.9%.

  • Company websites44.71%
  • Editorial media20.95%
  • Directories and databases12.95%
  • Reviews, forums, social, video11.39%
  • Reference and research6.99%
  • SEO filler and residue3.01%

Ranked lists, the "best X" pages, are 8.05% of all citations. 99.6% of those come from a third party with no stake in the outcome.

Does the assistant matter?

The assistant your buyer opens decides if your press coverage matters.

We scanned 19 brands on both Gemini and ChatGPT in the same month, so the brand mix cannot drive the gap.

Gemini drew on editorial media 21.06% of the time. ChatGPT did it 11.03% of the time. Gemini was higher for 17 of those 19 brands.

Median gap 10.51 points, p < 0.001. The coverage did not change and no outlet got better. The models read different things.

We see the same direction for Gemini against Grok on 10 brands, 24.89% to 16.93%, p = 0.006. Grok against ChatGPT is not in the claim: p = 0.055 at 8 brands.

0%20%40%60%GeminiChatGPT

Each line is one brand. It shows the media share of citations on each assistant. Green shows where Gemini is higher.

Where does criticism come from?

Almost everything critical an AI says about a brand arrived through journalism.

We tag sentiment only when the answer discusses the citation. That happened for 24,816 citations. Split the negative ones by source and the contrast is clear.

Media is 20.95% of the corpus and 42.03% of everything negative in it, about twice its weight.

Rewriting your own pages edits material that is 0.62% negative. The pages that make an answer critical belong to somebody else.

Share of citations carrying negative sentiment

National news39.37%
Forums18.33%
Trade press14.29%
Regional news13.25%
Blogs7.11%
Consumer press6.11%
Review sites2.21%
Press releases0.88%
Company websites0.62%

What happens in other languages?

Ask in another language and the model quotes a different press corps.

English is 49.3% of citations; the rest is 50.7%, across 16 languages. Local answers lean harder on media: Polish 28.99%, Estonian 27.37%, English 17.71%.

Paired within brand on 26 brands with enough of both, local is higher for 19 of them, p = 0.002.

For the same brand in the same scan, the outlets cited in English and the outlets cited in the local language overlap by a median 20.7%.

Weighting by volume reads the other way. 75.51% of non-English media citations land on a domain that also appears in the English answers. The head of each list is local. Most of the volume is shared. Plan on both numbers.

Polish media by brand reach, 9 brands

wirtualnemedia.pl8 of 9
press.pl7 of 9
proto.pl6 of 9
rp.pl6 of 9
mycompanypolska.pl6 of 9
pap.pl6 of 9
aboutmarketing.pl6 of 9

How many of the 9 Polish-language brands cite each outlet. Trade and PR press outrank the mass consumer titles.

How the question changes the answer

A prompt asking for a recommendation reads the press more than a prompt checking a name.

Answer Trail runs two waves in one session. Wave 1 asks what the market recommends. Wave 2 checks whether each named brand holds up. Same run, same engine, same day, same classifier. Only the question changes.

Carried by Grok, 78 runs at p = 0.002. Same direction and a larger gap on Gemini, 37 runs at p = 0.084. We did not test ChatGPT at 10 runs. These are cold-outreach checks on companies with no relationship to us. We hash subjects at collection and never write them to a table.

Media share of the sources each wave cited, mean across 125 paired runs

25.97%Wave 1, what do you recommend
20.97%Wave 2, does this brand hold up

Wave 1 was higher in 72 of 125 runs across 75 brands. Median gap 2.86 points, p < 0.001.

Who dominates the citations?

There is no shortlist. The top of the list barely moves.

The 50 largest domains hold 26.21% of citations here. In our published May index, built on 66 different brands three months earlier with different models, the 50 largest held 24.5%.

80.82%cited for exactly one brandindex corpus, 11,681 domains
80.12%cited for exactly one subjectAnswer Trail, 3,285 domains

Behind that stable head the corpus is almost all singletons. The two instruments land within a point of each other on brand sets that share nothing. Concentration comes from how the models source.

Sources by brand reach, 44 brands

wikipedia.org39 of 44
youtube.com38 of 44
reddit.com37 of 44
trustpilot.com34 of 44
indeed.com34 of 44
tracxn.com33 of 44
pitchbook.com32 of 44
forbes.com25 of 44

The default reading list for a B2B brand is an encyclopedia, a video site, a forum, two review sites and two funding databases. The first news outlet appears at rank 8.

What to do with this

Both conclusions are about where you look.

If you run a brand

An AI visibility number without a model name is an incomplete sentence. The same coverage is worth about twice as much in one assistant as in another.

If you are reading an AI answer

Most of the reading list is company pages and 20.95% of it is press. What you get depends on the assistant you opened and the language you asked in.

Method

How this was measured, and what it cannot tell you

  1. We counted 119,055 citations from 78 production scans of 44 brands, 13 July 2026 to 12 August 2026, in 16 languages. Gemini ran 46 of those scans, ChatGPT 20, Grok 11, and Claude 1.
  2. Source types come from a 24-value classifier. Media means national news, regional news, trade press, consumer press, blogs and wire releases together. Domains are resolved to the registrable domain, and a brand's citations of itself are dropped from every ranking.
  3. Engine and language comparisons are paired within brand and tested with a Wilcoxon signed-rank test. An unpaired comparison across this corpus would be reading the brand mix, because the models were not run on the same brand sets.
  4. Brands are anonymised and the corpus is aggregated. No customer is named, and every domain named in this deck is cited for at least 3 different brands.

What it cannot tell you: a per-country ranking, because the country field was wrong for 12.8% of the 39 brands hand-checked against their real headquarters; a per-industry ranking, because only 3 of 10 industry buckets reach 5 brands. Language carries 100% coverage and is the axis that worked. The single Claude scan is excluded from every engine claim.

The sample leans toward B2B software and services, so a market census would read differently. Every figure on these slides is computed from the study's results file at build time.

Try it now

See what AI cites about your brand.

The same check we ran across 44 brands takes about a minute to run on yours.

Check your brand →