When someone asks an assistant about a brand, or asks it to recommend one, the answer is built from what the model learned in training and what it reads while answering. Brand signals are the pieces of that material: who mentions you, what reviewers and list writers say, what Wikipedia and the knowledge graphs hold, how consistently your details are recorded, and what your own site says. What assistants make of them reaches more people each month. In Ahrefs' crawls of US desktop Google results, AI Overviews appeared on a monthly average of 61.3% of branded keywords in July 2026 and 73.3% in September, and on 82.91% of them on 29 September.1 An AI summary is now often the first description of a brand that a searcher reads.
This post covers the signals themselves: what each one is, how it reaches an answer and what the evidence shows. Correcting a wrong fact is covered in when an AI assistant gets your business wrong, measurement in how to measure visibility in AI search, and the full programme of work in the tutorial on getting recommended by AI assistants. The mechanics of retrieval are in the AI search guide.
Two routes from a brand signal into an answer
Training is the first route. Every model has a cutoff: Anthropic's latest Claude models were trained on data up to June 2026 and may not know about anything that happened later.2 What a model knows about a brand depends heavily on how often the brand appears in what it read. Kandpal and colleagues found that a model's accuracy on a fact-based question tracks the number of pre-training documents related to it, and that retrieval reduces that dependence.3 A brand written about on thousands of pages is likely to be well represented in a model. One mentioned on a handful of pages is long-tail knowledge, which models learn poorly.
Live retrieval is the second. Claude searches when a request depends on current or changing information, including details about specific organisations, people or products that might have changed.4 Google's AI Mode draws on sources such as the Knowledge Graph, information about the real world and shopping data, and issues several related searches at once across subtopics.5 Retrieval is where a smaller brand has the best chance of being described accurately, and the only route on which a change made this month can show up this month.
| Signal | How it reaches an answer | What the evidence shows |
|---|---|---|
| Mentions across the web | Training, through how often a brand appears, and retrieval, through the pages that mention it | Mentions in Reddit threads, on webpages and in news articles correlated with AI visibility at 0.67 to 0.75 across 10,000 US brands.6 |
| Third-party lists and comparisons | Mostly retrieval, because assistants search for "best" and "review" pages | "Best X" blog lists were 43.8% of the page types ChatGPT cited for recommendation prompts.7 |
| Reviews | Retrieval, from review sites and from summaries of public reviews | ChatGPT shopping results summarise reviews from public websites.8 Review sites are among the most cited sources in some assistants.9 |
| Wikipedia and Wikidata | Training, where published data mixes weight Wikipedia up,10 and retrieval | Wikipedia was cited in 41 of 57 branded AI Overviews in a 100-brand panel.1 |
| Consistent entity data | Retrieval and knowledge graphs, which have to match each record to one entity | Bing asks for entity names to be clear and consistent, with no ambiguous references.11 |
| Your own site | Retrieval for facts, and training | Official sources were cited for 84.4% of factual questions about pricing, product features or company details in ChatGPT.12 |
Mentions across the web
The most recent large study was published by Ahrefs on 3 November 2026. It compared seven kinds of off-site mention with how often 10,000 US brands were named across 260 million non-branded prompts in AI Overviews, AI Mode and ChatGPT, using Spearman correlation. Mentions in Reddit threads that appear in Google results correlated at 0.75, mentions on webpages and across domains at 0.70, news articles at 0.67, YouTube videos at 0.64 and TikTok at 0.60. Press coverage carried no extra weight over mentions on ordinary webpages, and ChatGPT was the least tied to off-site mentions of the three.6
It comes with cautions. Big brands get more mentions and more AI visibility, so size drives some of the pattern: within bands of brands with similar search volume, the correlation fell to between 0.34 and 0.54, which Ahrefs reads as mentions still predicting visibility independently of how well known a brand already is. I would treat the smaller figures as the fairer guide for a brand trying to grow. Reddit's lead is partly by design, Ahrefs notes, because Reddit licenses its content to Google and OpenAI.6 Ahrefs also sells the tools that track these mentions.
Earlier studies point the same way. Ahrefs' December 2025 study of 75,000 brands found branded web mentions correlated with AI visibility at 0.656 to 0.709 across the same three surfaces, while the number of backlinks and URL rating showed very weak correlations.13 In January 2025, Seer Interactive reported running 10,000 finance and SaaS questions through the GPT-4o API, and found brand mentions correlated at about 0.65 with page-one Google rankings and about 0.5 to 0.6 with Bing rankings, with backlinks showing weak or neutral effects.14 None of these studies can show that a new mention causes a new recommendation, and each presents its results as correlations.
The two routes explain why mentions would matter. In training, the number of documents that mention an entity shapes how well a model can answer questions about it.3 In retrieval, the pages that mention you are the pages a search can find, and Google says its AI features can show what is being said about products and services across the web.15 The links and digital PR guide covers how coverage is earned.
Third-party lists and comparisons
When a prompt asks for a recommendation, the sources are often other people's lists. In a study published in December 2025, Glen Allsopp at Ahrefs analysed 26,283 source URLs cited for 750 top-of-funnel ChatGPT prompts asking for software, products and agencies. "Best X" blog lists made up 43.8% of the page types cited, 79.1% of the 1,100 lists with clear dates were last updated in 2025, and brands placed higher in a list appeared more likely to be recommended. In software, a self-promotional list, published by one of the brands it ranks, still showed up in more than a third of responses. For products it was rare.7 Ahrefs publishes lists of this kind itself and sells AI visibility tools.
An academic study found the same lean towards third parties. Mahe Chen and colleagues at the University of Toronto sent ranking-style prompts, such as requests for the top ten brands in a category, to Perplexity, Claude, Gemini and GPT-4o search in August 2025, and compared the domains they cited with Google's top ten. They sorted sources into brand-owned sites, social platforms and earned media, meaning independent media, review and comparison sites. For US consumer electronics, 92.1% of the AI citations were earned media. For US software, 43.7% of Google's results were brand sites against 26.7% of the AI citations. ChatGPT was among the most earned-heavy engines, at 93.5% earned for well-known brands, and Gemini the most brand-leaning, at 25.1% brand.16
The same paper found a lean towards market leaders. Across 50 unbranded cola prompts on ChatGPT and Perplexity, 62.2% of the brands named were major brands and 9.0% niche ones. The authors stress that the data is a snapshot from August 2025, that the lines between their categories involve judgement, and that the exact shares are best read as signs of relative trends.16 It is an arXiv preprint.
Assistants find these pages partly because they search for them. In logged-out ChatGPT responses from August and September 2026, Ahrefs found "best" in 5.3% of September's fan-out queries, "reviews" in 3.6% and the year "2026" in 22.9%.12 In my view a recently updated list that names you, with accurate details, is one of the more direct signals a brand can earn. The SEO power of listicles in the age of ChatGPT covers lists in more depth.
Reviews
Reviews are the raw material for a common brand question, whether a business is any good. OpenAI says ChatGPT's shopping results can show review summaries that the model writes from reviews on public websites, and that OpenAI does not verify those reviews or ratings. It ranks merchants on factors such as availability, price, quality and whether the merchant makes or is the primary seller of the item.8 For local businesses, reviews are also a documented ranking signal: Google says more reviews and positive ratings can help local ranking, and that prominence is partly based on how many reviews a business has.17
Review sites are also sources in their own right. In Profound's analysis of 680 million citations from August 2024 to June 2025, Yelp and Tripadvisor were among Perplexity's ten most cited sources and G2 among ChatGPT's.9 SE Ranking modelled ChatGPT citations across 216,524 pages on 129,000 domains in 20 niches, published in November 2025, and found that domains with profiles on platforms such as Trustpilot, G2, Capterra, Sitejabber and Yelp had three times the chance of being cited as a source.18 That is an association from a predictive model, and it cannot say which came first.
Ahrefs' tracking of about 37,500 citations across five assistants over 20 days, published in October 2026, adds a small result worth knowing. For questions about Ahrefs' own pricing, AI Overviews and AI Mode cited Ahrefs' pages on 97% to 100% of days. For review questions about Ahrefs, they cited them on 0% to 5% of days.19 It covers 22 questions about one brand, so treat it as an example. It fits the wider pattern: for an opinion about you, assistants look to someone other than you. The rules on asking for reviews, and the UK law on fake and incentivised ones, are in the post on review incentives and review gating.
Wikipedia and Wikidata
Wikipedia reaches assistants strongly by both routes. In the training mixes that labs have published, it is sampled far more heavily than its size alone would give it. GPT-3's mix in 2020 gave Wikipedia's 3 billion tokens a 3% weight, so it was seen 3.4 times over the 300 billion training tokens, against 0.44 times for filtered Common Crawl, and the authors say datasets they viewed as higher quality were sampled more often.10 Meta's LLaMA in 2023 used Wikipedia dumps from June to August 2022 in 20 languages at 4.5% of the mix, seen 2.45 times.20 The labs behind today's leading assistants do not publish their mixes, so whether Wikipedia keeps that weight is an inference.
In retrieval the picture moves. In Ahrefs' 100-brand panel of AI Overviews for bare brand names, Wikipedia was cited in 41 of the 57 answers that showed sources, YouTube in 22 and LinkedIn in 13.1 Profound found Wikipedia was ChatGPT's most cited source between August 2024 and June 2025, at 7.8% of all its citations.9 By 2026 that had changed. In Ahrefs' tracking of US ChatGPT answers, Wikipedia went from being cited in 1 in 7 answers in March to 1 in 35 by late September, and Reddit from 1 in 5 to 1 in 43.12 Any figure for a source's share of citations is a snapshot.
Google's knowledge panels are generated automatically from information across the web and update as it changes, and the subject of a panel can claim it and suggest changes.21 Wikidata accepts an item that has a Wikipedia sitelink, or that is a clearly identifiable entity which can be described using serious, publicly available references.22 Schema.org gives a Wikipedia page and a Wikidata entry as examples of what sameAs should point to.23 Entity optimisation covers creating and checking a Wikidata item.
A Wikipedia article is something a brand earns through coverage. A company is presumed notable when it has had significant coverage in multiple reliable secondary sources that are independent of it, and press releases and routine announcements do not count.24 People with a conflict of interest are strongly discouraged from editing the article directly and are asked to propose changes on its talk page, and paid editors must disclose who is paying them.25 If a brand is not notable yet, the work is the coverage.
Consistent entity data
An assistant has to work out that the brand on your site, your Business Profile, a review site and a news story is one entity. Bing's guidelines, rewritten in February 2026, say entity names should be clear and consistent with no ambiguous references, as part of their guidance on grounding AI answers.11 Organization markup can help Google understand an organisation's administrative details and disambiguate it in search results, with sameAs pointing to pages elsewhere that say more about it.26 Google's page on AI features lists checking that Merchant Center and Business Profile information is up to date as a best practice.27
Inconsistent naming even trips up the people measuring. In Ahrefs' citation tracking, a manual check of 55 brands found that 30% to 36% of apparent losses were the same brand under a different name.19 The practical work of one name and one set of details everywhere is in entity optimisation, and tracing a wrong fact to its source is in the post on wrong answers.
Your own site
A brand's own site is cited less than third parties when the prompt asks for a ranking, and more when it asks for facts. In the Toronto study, the AI engines cited brand-owned sites less often than Google did.16 In Ahrefs' ChatGPT study, 84.4% of factual questions about pricing, product features or company details cited at least one official source, against 68.3% of comparison questions, and 63.4% of citations for factual questions went to official websites. Of the most common site: searches ChatGPT ran, 75.9% were restricted to brand sites such as apple.com or ikea.com.12 When a question named the brand, its own site was cited on all 20 days in roughly 28% to 33% of cases on Google's tools and ChatGPT.19
So the site is where an assistant checks what you charge, what you offer and who you are, and how the page is written matters. The GEO paper by Aggarwal and colleagues, accepted at KDD 2024, tested nine ways of rewriting source pages on a benchmark of 10,000 queries, on an answer engine the authors built with GPT-3.5 and on Perplexity. Adding citations to sources, quotations and statistics raised a page's visibility in the answer by about 30% to 40% on their main measure, which counts the words attributed to a source and weights them by position. Keyword stuffing scored below the unchanged page, and lower-ranked sources gained more: the fifth-ranked source's visibility rose 115.1% with cited sources added.28 That was a research setup in 2023 and 2024, so I read it as support for writing specific, sourced facts that can be quoted, without expecting gains of that size. Markup is covered in does schema markup help AI citations.
What the studies cannot tell you
Every figure above is a correlation or an outside-in observation. No assistant provider publishes how it weights sources, and Google says no third-party tool has access to its internal ranking or AI systems.15 Most of the industry studies come from companies that sell AI visibility tracking, the samples are mostly from the US, and the citation mix shifts month to month, as the 2026 falls in Wikipedia and Reddit citations show.12 Answers also change from day to day. In Ahrefs' 20-day tracking, 59% of ChatGPT's citations were gone by the second day, while the brand names in answers were steadier than the pages cited.19
Recommendations can be gamed, and the platforms now address it directly. Kumar and Lakkaraju showed in 2024 that adding a crafted text sequence to a product's page in a catalogue of fictitious coffee machines significantly increased the chance that a language model listed it as its top recommendation, and warned that this could disrupt fair market competition.29 Bing's February 2026 guidelines added a section on prompt injection and AI manipulation, and renamed the keyword stuffing section to cover artificially engineered language, including content designed to trigger citations or AI responses.11 Google says its AI features depend on its spam systems as well as its ranking systems, that seeking inauthentic mentions across the web is not as helpful as it might seem, and that manipulating generative AI responses in Search violates its spam policies.15
How to strengthen brand signals honestly
The honest version of this work is slower, and it holds up when the platforms change. This is the order I would work in.
- Make your own site the reference for your facts. Prices or how pricing works, what you offer, who you suit, where you operate, your policies and your credentials, in plain text on crawlable pages. These are the questions assistants answer from official sources.12
- Use one name and one set of details everywhere, with Organization markup and sameAs links to the profiles that are unmistakably yours.26
- Earn mentions where your customers already read and talk: trade press, specialist sites, communities and video. In Ahrefs' data, press coverage carried no more weight than mentions on ordinary webpages, and weighting mentions by the traffic of the page made no difference to the correlation.6
- Take part in Reddit and forums as yourself, answering questions in your field and saying who you are. Reddit mentions were the strongest correlate of AI visibility in the same study.6
- Get into the lists that already exist. Find the "best" lists your test prompts cite and send their editors accurate, current details. If you publish your own comparison, keep it fair and checkable. Under the Digital Markets, Competition and Consumers Act 2024, paying for editorial content that promotes you without making the payment clear is banned in the UK,30 and the advertising code requires marketing to be obviously identifiable as marketing.31
- Ask every customer for an honest review on the platforms your sector uses. Fake reviews and reviews that conceal an incentive are banned in the UK under the same Act.3032
- Let Wikipedia follow coverage. If the brand meets the notability bar, disclose your connection and propose changes on the talk page.25
- Skip hidden text for language models, instructions injected into pages and bought mentions. Bing's guidelines and Google's spam policies both cover attempts to manipulate AI answers.1115
Ways to build brand signals, from fine to unlawful
Earning coverage in trade press, specialist sites and communities
Mentions across the web are the strongest measured correlate of being named by assistants.
Fine to doFine to doAsking every customer for an honest review
Assistants summarise public reviews, and review sites are among the sources they cite.
Fine to doFine to doPublishing your own "best" list or comparison page
It shows up in ChatGPT sources for software, but it is your own marketing, so it has to be fair and checkable.
A grey area: handle with careA grey area: handle with careEditing the Wikipedia article about your own company
Wikipedia strongly discourages it and asks for proposed changes on the talk page, with the connection disclosed.
A grey area: handle with careA grey area: handle with careBuying mentions or hiding text to steer AI answers
Manipulating generative AI responses violates Google's spam policies, and Bing's guidelines cover it too.
Against Google's policiesAgainst Google's policiesPaying for editorial coverage or a list placement without saying so
Banned under Schedule 20 to the Digital Markets, Competition and Consumers Act 2024.
Against the law (UK)Against the law (UK)Fake reviews, or reviews that hide an incentive
Banned under paragraph 13 of the same schedule to the Act.
Against the law (UK)Against the law (UK)
How to check what assistants say about a brand
Checking is a sampling job, set out in full in how to measure visibility in AI search. For brand signals the useful split is by prompt type, because each type leans on different signals.
| Prompt type | Example | Signals it mostly tests |
|---|---|---|
| Brand | "What is Fernbrook Tea?" or "Is Fernbrook Tea any good?" | Your own site for facts; reviews, Wikipedia and entity data for the description |
| Category | "Best loose-leaf tea shops in the UK" | Mentions, third-party lists and reviews |
| Comparison | "Fernbrook Tea vs [a competitor]" or "Alternatives to Fernbrook Tea" | Comparison pages and reviews, checked against your own facts |
Map the sources that describe a brand
Fill in the parts in brackets before you send it
Tools package this as a score. Semrush's AI Visibility score runs from 0 to 100 and combines how many topics include your brand in AI answers with how often you are mentioned within them, drawn from a database of more than 317 million prompts and responses.33 Peec AI's Visibility Score is the share of tracked responses that mention your brand,34 and it separates brand visibility, where you are named, from source visibility, where your content is used or cited without your name.35 Each vendor defines its score differently, so compare a brand with itself and its competitors in one tool, over months.
The signals that move assistants are mostly the ones that already describe a good business accurately in many places. The work is making sure that is true, and that it is written down where an assistant can find it.
Common questions
01What are brand signals in AI search?
02What signals make AI search engines trust a brand enough to recommend it?
03What are the most common signals AIs use to describe brands?
04What is brand entity optimisation?
05How do you measure brand visibility in AI search?
Sources
- Platform docs15
- Regulator4
- Industry study9
- Research6
- Reporting1
- 01AI Overviews Currently Appear for 83% of Branded SearchesAhrefs (Ryan Law, October 2026)Industry study
- 02How up-to-date is Claude's training data?Claude Help Center (Anthropic)Platform docs
- 03Large Language Models Struggle to Learn Long-Tail KnowledgeKandpal, Deng, Roberts, Wallace and Raffel, ICML 2023 (arXiv)Research
- 04Web search toolClaude Platform Docs (Anthropic)Platform docs
- 05Expanding AI Overviews and introducing AI ModeGoogle (Robby Stein, March 2025)Platform docs
- 06Which Types of Mentions Correlate the Strongest With AI Visibility? (260M AI Prompts Analyzed)Ahrefs (Louise Linehan, November 2026)Industry study
- 07Do Self-Promotional “Best” Lists Boost ChatGPT Visibility? Study of 26,283 Source URLsAhrefs (Glen Allsopp, December 2025)Industry study
- 08Shopping with ChatGPT SearchOpenAI Help CenterPlatform docs
- 09AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source InformationProfound (Nick Lafferty, June 2025, updated August 2025)Industry study
- 10Language Models are Few-Shot LearnersBrown et al., OpenAI, arXiv (May 2020)Research
- 11Bing Adds GEO To Official Guidelines, Expands AI Abuse DefinitionsSearch Engine Journal (Matt G. Southern, February 2026)Reporting
- 12How ChatGPT Does Query Fan-Out (15 Million Queries Studied)Ahrefs (Louise Linehan, October 2026)Industry study
- 13Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)Ahrefs (Louise Linehan, December 2025)Industry study
- 14STUDY: What Drives Brand Mentions in AI Answers?Seer Interactive (Christina Blake and Nick Haigler, January 2025)Industry study
- 15Optimizing your website for generative AI features on Google SearchGoogle Search CentralPlatform docs
- 16Generative Engine Optimization: How to Dominate AI SearchChen, Wang, Chen and Koudas, University of Toronto, arXiv (September 2025)Research
- 17Tips to improve your local ranking on GoogleGoogle Business Profile HelpPlatform docs
- 18How to optimize for ChatGPT: LLMs.txt doesn’t matter but brand mentions on Quora and Reddit doSE Ranking (Yulia Deda, November 2025)Industry study
- 19How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 CitationsAhrefs (Mateusz Makosiewicz, October 2026)Industry study
- 20LLaMA: Open and Efficient Foundation Language ModelsTouvron et al., Meta AI, arXiv (February 2023)Research
- 21About knowledge panelsGoogle Search HelpPlatform docs
- 22Wikidata:NotabilityWikidataPlatform docs
- 23sameAsSchema.orgRegulator
- 24Wikipedia:Notability (organizations and companies)WikipediaPlatform docs
- 25Wikipedia:Conflict of interestWikipediaPlatform docs
- 26Organization (Organization) structured dataGoogle Search CentralPlatform docs
- 27AI features and your websiteGoogle Search CentralPlatform docs
- 28GEO: Generative Engine OptimizationAggarwal et al., KDD 2024 (arXiv)Research
- 29Manipulating Large Language Models to Increase Product VisibilityKumar and Lakkaraju, arXiv (April 2024)Research
- 30Digital Markets, Competition and Consumers Act 2024, Schedule 20legislation.gov.ukRegulator
- 31Recognising ads: Advertisement featuresAdvertising Standards Authority / CAPRegulator
- 32Short guide for businesses: publishing consumer reviews and complying with consumer protection lawCompetition and Markets Authority (April 2025)Regulator
- 33Where does the data in Semrush’s AI Visibility Toolkit come from?Semrush Knowledge BasePlatform docs
- 34VisibilityPeec AI DocsPlatform docs
- 35Understanding your metricsPeec AI DocsPlatform docs
Written by Mani Bharij, SEO & AI Search Consultant in London. More on this subject in the AI search guide.