Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 13 min

AI search cites what others say about you

Generative engines mostly cite what other sites say about a company, not what the company says about itself. In a June 2026 study of 167,551 citations about 128 brands, 85.7% pointed to sites the brand did not own. A company that ranks well on Google can still be missing from ChatGPT or Perplexity answers, because those answers draw on reviews, comparison pages, encyclopedias, forums, and press. The work is to earn accurate coverage on the sources engines already retrieve, and to stop chasing shortcuts that platforms now punish.

In short

  • Most citations about a brand are third-party. One study put owned-site citations at 14.3% of the total. Your own site still matters, but it is the minority of what engines quote.
  • Brand mentions correlate with AI visibility more strongly than backlinks do. Ahrefs found branded web mentions correlated at about 0.66 with AI Overview visibility, and backlinks at 0.22. That is correlation, not proof that mentions cause citations.
  • The sources that show up for B2B queries are review sites, third-party comparison lists, LinkedIn, Wikipedia, YouTube, trade press, and Reddit. Their weight differs by engine and moves month to month.
  • Reddit’s share of ChatGPT Search citations fell by about 86% in mid-August 2026 in one vendor’s data, and Reddit is actively removing covert brand posts. “Seed Reddit threads” is now stale advice and a ban risk.
  • Wikipedia is earned, not written. Its notability rule requires independent coverage, and its conflict-of-interest rule tells company staff not to edit their own article.
  • Measure it as a source gap: for each query you care about, list the domains engines cite where you are absent, and work on the ones you can legitimately influence.

Why do AI engines cite third-party sites instead of your own?

They cite third parties because an answer about a company is more credible when someone other than the company says it, and because retrieval finds far more pages about you than by you. The clearest number comes from an arXiv paper by Dmitrij Zatuchin, submitted June 24, 2026, which merged three Rankfor.AI datasets covering 128 brands across 12 home markets and 13 languages. Of 167,551 URL-grounded citations, 85.7% pointed to sites the brand does not own and 14.3% to owned sites. Wikipedia was the most-cited domain in 11 of the 12 languages. The paper does not name which engines produced the citations, so read it as a pattern, not a per-engine rule.

A Reddit commenter put the buyer-side version plainly, in paraphrase: Google ranks what you publish, and AI search surfaces what people say about you. That is half right. Google also weighs what others say, and AI engines also read your site. But the balance has shifted. On a classic results page your product page can hold position one. In a generated answer it is one source among several, often next to a review site and a comparison article, and the model is looking for agreement between them.

This is the corroboration property in what gets cited: a claim that independent sources confirm is easier for a model to reuse than one that exists only on the vendor’s page. It is also why GEO cannot be done entirely on your own domain.

They correlate more strongly, and that is all the evidence shows. Ahrefs’ May 2025 study of about 75,000 brands measured Spearman correlations between site and brand factors and how often each brand appeared in AI Overviews:

FactorCorrelation with AI Overview brand visibility
Branded web mentions0.664
Branded anchors0.527
Branded search volume0.392
Domain Rating0.326
Number of backlinks0.218

Ahrefs’ December 2025 follow-up extended the analysis to ChatGPT and AI Mode. Branded web mentions correlated at 0.664 with ChatGPT visibility, 0.709 with AI Mode, and 0.656 with AI Overviews. YouTube mentions were the strongest factor, at about 0.737 across the three. Backlinks were described as very weak, and the number of pages on a site correlated at about 0.19.

Three cautions belong next to that table.

  1. Correlation is not causation. Ahrefs says so in both studies, and adds that these factors may simply be “markers of existing brand strength.” Large, well-known brands score high on mentions, search volume, and AI visibility at the same time. Buying mentions does not make you a large, well-known brand.
  2. The sample is filtered. Brands needed a Domain Rating above 40 and a top keyword with at least 800 monthly searches, and about 26% of brands in the original sample had no AI Overview mentions at all. Small or new companies are underrepresented.
  3. The figures come from one vendor’s tool. They describe Ahrefs’ tracked prompts and its definition of a mention.

The practical reading is narrower than the headlines that followed. Backlinks still matter for retrieval in classic search, which feeds several AI systems. But a link campaign that does not produce people writing about you in their own words is unlikely to move AI answers, and a company that is discussed widely tends to be named in them.

Which third-party sources show up for B2B queries?

Seven source types recur across studies of B2B and software queries. None of them is a fixed winner, and the evidence for each varies in quality.

Source typeWhat the evidence saysStrength of evidenceWhat you can legitimately do
Review sites (G2, Capterra, GetApp, TrustRadius)Frequently cited for software queries, but rankings conflict by study and engine. A G2-published analysis of about 30,000 citations found a 10% increase in reviews associated with roughly 2% more citations, explaining under 2% of the variationWeak to moderate, several studies by interested partiesKeep profiles complete and current; ask real customers to review
Third-party comparison lists (“best X for Y”)Peec’s analysis of nearly 200,000 responses found rank #1 in a frequently cited third-party list associated with a 16.5-point visibility lift in B2B SaaSModerate, observational, vendor studyPitch editors with evidence; correct factual errors on lists that cover you
LinkedInProfound reported LinkedIn as the most-cited domain for professional queries on six AI platforms, Nov 2025 to Feb 2026, with posts and articles gaining share over profilesModerate, vendor study, synthetic prompts for the top-line claimPublish substantive posts and articles from named experts
WikipediaMost-cited domain in 11 of 12 languages in the arXiv studyModerateEarn coverage that meets notability; never edit your own article directly
YouTubeStrongest correlate of AI visibility in Ahrefs’ December 2025 data (about 0.74)Correlation onlyPublish real explainers and demos with accurate titles and transcripts
Trade press and analystsProfound’s August 2026 analysis of 158,000 claims found inaccurate claims came mostly from competitor and earned-media pages, so press coverage shapes what engines repeatIndirectEarned media built on data or expertise
Reddit and forumsLarge but volatile share; sharply reduced in ChatGPT in August 2026Contested, tracker-dependentParticipate openly as yourself, or not at all

Two notes on the review-site row. The G2 study was written by Kevin Indig and published on G2’s own blog, and he calls the results “suggestive associations rather than causal effects.” Other trackers rank GetApp or Capterra above G2 in some engines. The safe conclusion is that review sites are a common source for software shortlists, and that which one matters for your category is something to measure, not assume.

On comparison lists, Peec’s study (May 14, 2026, by Jan Ehrlinspiel) covered eight engines across B2B SaaS, emerging martech, and US finance, from September 2025 to March 2026. It restricted the analysis to third-party lists that engines repeatedly cite, and found that a few placements in those lists outweighed many placements in lists engines never retrieve. It did not test lists a vendor writes about itself. Peec sells AI monitoring tools, and the study is observational. The lesson that holds up: one accurate entry in a list engines actually read is worth more than fifty entries in lists they ignore.

What happened to Reddit citations in August 2026?

Reddit’s share of ChatGPT Search citations dropped sharply in mid-August 2026 in Promptwatch’s tracking data, and the cause has not been confirmed. Promptwatch’s August 18 report put Reddit at an average 3.83% of ChatGPT Search citations from July 18 to August 7. The share slid on August 8 and fell below 1% on August 14, averaging 0.52% from August 14 to 17, an 86.4% relative drop. Over the same window, Reddit’s share fell about 11% in Google AI Overviews and about 31% in AI Mode, both more gradually.

Read the number with its caveats. Promptwatch itself called the size of the drop provisional and said a data-collection issue could not be ruled out. Search Engine Journal noted that the popular explanation, an August 8 change in which ChatGPT began issuing many more site: queries during fan-out, does not match the sharper August 14 step. OpenAI did not explain the change. Baselines also differ by tracker, and SEJ advised checking your own domain’s citations before changing strategy on one vendor’s average. A similar drop happened in September 2025. The drop measures citations in one vendor’s panel. It does not show that ChatGPT stopped using Reddit content in other ways, and the AI search statistics guide explains why trackers disagree this much.

The durable lesson is volatility. If a third of a program’s AI visibility rested on one platform’s citation share, that program lost most of it in a week through no action of its own.

Why is “seed Reddit threads” advice stale and risky?

It is stale because the citation share it was chasing has shrunk and moves without warning, and risky because Reddit is now hunting the practice. Between 2024 and mid-2026, a common GEO tactic was to post or pay for brand-friendly Reddit comments so that AI engines would pick them up. Reddit’s response, described in its July 6, 2026 post on keeping Reddit “real,” and reported by Engadget, includes:

  • About 25,000 spammy posts and comments caught per day, and close to 2 million inauthentic votes revoked per day.
  • Large language models used at account creation and afterward to look for “highly subtle, coordinated patterns of fake behavior and artificial hype.”
  • Human verification for accounts flagged as possibly automated.

The costs of being caught are concrete: removed posts, banned accounts, and threads where the top comment is a user pointing out that your company is posting under fake names. That thread can then be the one an AI engine retrieves.

What still works on Reddit is what always did. Answer questions in your area under an account that says who you are, follow each subreddit’s self-promotion rules, and fix factual errors in threads about you with a disclosed reply. If your customers already discuss you there, that discussion is a source you can learn from. It is not a channel you can manufacture.

Can you get your company onto Wikipedia?

Only by becoming notable through independent coverage, and not by writing the article yourself. Wikipedia’s notability guideline for organizations says a company is presumed notable when it has “significant coverage in multiple reliable secondary sources that are independent of the subject.” It excludes press releases, the company’s own website, “trivial or incidental coverage,” and “routine coverage” such as funding announcements.

Its conflict-of-interest guideline is explicit about company staff and agents: “COI editors are strongly discouraged from editing affected articles directly.” They should disclose the connection and propose changes on the article’s talk page through an edit request, so independent editors can review them. Uncontroversial fixes such as reverting vandalism or correcting a spelling error are allowed. Paid editors must, under the Wikimedia Foundation’s Terms of Use, “disclose your employer, client, and affiliation.”

So the honest playbook is short:

  1. Do not pay anyone to write or “manage” your article. Undisclosed paid editing breaks the Terms of Use, and a disclosed one still has to pass independent review.
  2. If an article exists and is wrong, request the correction on its talk page. Disclose who you are and cite an independent source for the correct fact.
  3. If no article exists, work on the precondition. Earn the independent coverage that notability requires. An article will follow if and when editors judge that it should.

Wikipedia matters to AI answers beyond its own citations, because other sites copy its facts. A wrong fact there tends to propagate. The AI brand accuracy guide covers how to trace and correct that.

How do you legitimately earn third-party corroboration?

Give other people a reason and an easy way to describe you accurately. Four kinds of work do that, and none of them depends on a loophole.

  1. Publish citable data. A benchmark, a pricing index, a survey with a stated method, or a dataset others can reuse gives journalists and list writers something to cite with your name on it. The data has to be on an HTML page with a date and a method note, not only in a PDF.
  2. Keep analyst and review profiles accurate and active. Complete your profiles on the review sites your category actually uses, keep pricing tiers and feature lists current, and ask real customers for reviews through normal channels. Do not incentivize reviews in ways the platform prohibits.
  3. Make directory and entity profiles agree. Your company name, category, founding year, headquarters, product names, and pricing model should match across LinkedIn, Crunchbase, review sites, your Google knowledge panel, and your own site. The structured data guide covers how sameAs ties these together. The profiles have to be correct before the markup points at them.
  4. Earn media coverage on substance. Trade press, podcasts, and analyst notes that quote your people on a topic they know are the coverage notability rules and AI engines both reward. A press release that only you publish is not coverage.

A useful test for any tactic: would it still be worth doing if AI search did not exist? Accurate review profiles help buyers on G2. Original data earns links and press. Expert posts on LinkedIn build a following. Fake Reddit comments only make sense as an attempt to fool a model, and that is why they keep getting caught.

How do you find which third-party sources you are missing?

Run a source gap analysis: for each priority query, log which domains the engines cite, and mark the ones where competitors appear and you do not. The method follows the logging in measuring AI search visibility:

  1. Fix the query set. Use the commercial and comparative queries from your share of answer universe, reviewed by someone in sales.
  2. Run each query on a fixed engine panel, more than once. Answers vary run to run, so a single check will mislead you. Log the answer, every cited URL, and the date.
  3. Group citations by domain and source type. Review site, list article, publisher, forum, encyclopedia, video, vendor site.
  4. Mark presence. For each cited third-party URL, record whether it mentions you, mentions you accurately, or omits you while naming competitors.
  5. Rank the gaps by what you can influence. A list article that omits you and accepts updates is actionable. A Reddit thread from 2023 is not.

Worked example (illustrative)

The table below shows the shape of the output for an invented mid-size B2B payroll software vendor across 20 commercial queries on four engines, three runs each. The vendor, domains, and counts are illustrative, not client data.

Cited domainSource typeQueries where citedVendor mentionedCompetitors mentionedAction
A large review siteReview site14 of 20Yes, pricing outdatedYesUpdate tiers on profile
A trade publication’s “best payroll tools” listComparison list9 of 20No5 competitorsPitch the editor with the vendor’s published compliance data
A software comparison blogComparison list6 of 20Yes, wrong categoryYesRequest factual correction
Wikipedia, “Payroll software” articleEncyclopedia5 of 20No2 competitorsNone directly; work on independent coverage
A Reddit thread from 2024Forum4 of 20Yes, one negative commentYesNone; note as context for support
Vendor’s own pricing pageOwned3 of 20YesNoKeep; make sure HTML matches current pricing

The readout is not “get more mentions.” It is three specific actions on sources the engines already retrieve for these queries, two accuracy fixes, and two sources to leave alone. Re-run the same panel after the changes, and expect results to move with engine updates as well as with your work.

What are the common mistakes?

  • Treating owned content as the whole program. Most citations about a brand are third-party. Map them before writing another blog post.
  • Reading a correlation as a lever. Mentions correlate with visibility partly because big brands have both. Earn mentions that would be valuable on their own.
  • Building on one platform. Reddit’s ChatGPT citation share showed how fast a source can drop. Spread effort across source types.
  • Seeding forums under fake accounts. Reddit detects and bans it, and the exposure thread can become the cited source.
  • Paying for a Wikipedia article. Undisclosed, it breaks the Terms of Use. Disclosed, it still needs notability. Earn the coverage notability requires.
  • Counting mentions without checking them. A citation that states last year’s price is a liability. Score accuracy as its own number.
  • Trusting one tracker’s averages. Check your own domains and queries before changing strategy because of an industry chart.

Third-party presence is not a separate discipline from SEO, AEO, and GEO. It is the corroboration step in the same system: engines retrieve pages, extract claims, and cite the claims that independent sources agree on. If you want the source gap measured for your own query set, that is part of a GEO diagnostic.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit