Brand mentions in AI search come from five kinds of page: review and comparison directories, practitioner communities, trade and analyst coverage, partner and integration pages, and your own cited research. Assistants ground answers in publicly crawlable pages, so the fastest route to being named is to be described accurately on pages you do not control.
You already know mentions matter. The gap is almost always the next question: which pages, in what order, and who does the work. This piece is the earning motion rather than the argument. If you need the category framing first, it is in our explainer on what LLMO actually is. Everything below is a target list you can hand to a marketer on Monday.
The short answer: five source types carry most brand mentions in AI search
There is no published breakdown of AI mention share by source type for B2B categories, and anyone quoting one to two decimal places is guessing. What does exist is a measurement of which domains get linked. Pew Research Center's browsing data study, using browsing data from March 2025, found that Wikipedia, YouTube and Reddit were the most commonly linked sources in both Google AI summaries and standard search results, together accounting for 15 percent of the sources listed in the AI summaries examined, against 17 percent of sources in standard results. The concentration is real but modest, and the remaining majority is a long tail of ordinary web pages.
The taxonomy below is a working taxonomy drawn from our own client work, not a measured share. It is useful because each type has a different owner, a different cycle time and a different failure mode, which is what you need to sequence the work rather than admire it.
Why a mention on a page you do not own beats a page you do
Your own site says what you want it to say, and an assistant knows that. A third party page is evidence of the same claim from somebody with no incentive to flatter you, which is why it carries further when an answer has to characterise you in one sentence. We have made the full case for mentions against links in brand mentions versus backlinks in AI search, so this post assumes it rather than re-arguing it.
The mechanism is unglamorous. Google's documentation on AI features states that to appear as a supporting link in AI Overviews or AI Mode a page must be indexed and eligible to be shown with a snippet, with no additional technical requirements. Whichever page best answers the question and meets that bar can be used, and nothing reserves that slot for the vendor. This is also why unlinked brand mentions count here: the assistant is summarising the passage, not following the anchor.
One piece of housekeeping makes every later mention more useful. Use Google's Organization structured data on your home or about page. sameAs is for linking out to a page on another website with additional information about your organisation, while Google names iso6523 and naics as the properties used behind the scenes to disambiguate your organisation from others. If your brand name is ambiguous, mentions are getting attributed to somebody else and no amount of outreach fixes that.
Source type 1: review and comparison directories
Category directories are the densest concentration of comparative language about your market anywhere on the web, and their pages are built to be crawled. They answer the shortlist question directly, which is what your buyer is asking an assistant. That makes them the highest yield place to start, and the one place you can improve in an afternoon.
The work is maintenance, not campaigning. Claim every profile in your category. Fix the category selection, the one line description, the feature list and the integration list, because those strings are what gets summarised. Keep pricing consistent with your own site, since contradictory pricing is a common reason an assistant hedges, which is the same logic behind optimising pricing pages for AI search. Then build a standing motion for recent reviews from real customers, because recency is visible and staleness reads as decline. Never incentivise a review in a way the platform forbids, and never write one.
Source type 2: practitioner communities
Reddit is among the three most commonly linked domains in Pew's analysis, and the reason is that community threads contain the sentences buyers actually want: what broke, what it cost, what people switched to. Read this section as a durable motion rather than a current ranking of sources, because ChatGPT's own citation mix moved sharply away from community content in August 2026. That is also why this source type is the one most often ruined by the brand that tries to farm it.
The rules are simple and non negotiable. Participate as named humans with the affiliation stated. Answer the specific question asked, including when the honest answer is that a competitor fits better. Do not drop links, do not seed threads, do not run sockpuppets. This is the slowest of the five types and the hardest to fake, which is exactly why a mention here is worth more than five on pages you paid for. If your near term goal is to get mentioned by ChatGPT specifically, note that it cut community citations sharply in August 2026: Reddit fell from roughly 3.83 percent to 0.52 percent of ChatGPT citations between early and mid August 2026, with documentation, help centres and official brand pages taking the share. The B2B SaaS version of this motion is in our LLMO guide for B2B SaaS.
Source type 3: trade and analyst coverage
Trade press and analyst notes are what an assistant reaches for when it needs an authoritative characterisation rather than a user opinion. They are also the slowest to earn, because most pitches offer an announcement rather than a fact worth repeating.
Give the writer something checkable. A specific number with a stated method, a named expert with real credentials, and a claim that can be verified without taking your word for it. Google's guidance on creating helpful, reliable, people-first content frames the same test through its Who, How and Why questions: is it self evident who authored the content, does the page carry a byline, and does that byline lead to more about the author. Journalists apply that test informally and assistants surface the result. Transcribe your podcast appearances, because an audio file is not retrievable text. And accept that coverage will not rescue a weak product story, which is the point made in AEO will not fix average content.
Source type 4: partner and integration pages
This is the most neglected of the five and the cheapest to fix. Marketplace listings, integration directories and partner case studies already exist, already rank, and already describe you. They usually describe you badly, using copy somebody wrote at launch and nobody has touched since.
Run it as an audit. List every marketplace, app directory and partner site that carries a page about you. Check the description, the category, the screenshots and the stated capabilities against what the product does today. Send corrections. Then ask your best partners to write up the joint use case in their own words, because a partner describing the problem you solved together is a far stronger passage than a mutual logo swap. The way Clay compounds its own ecosystem pages is a useful model for what this looks like at scale.
Source type 5: your own cited research
Original research is the only source type on this list that you can manufacture, and it works by turning into somebody else's page. A number that other publishers quote travels with your name attached, which converts one asset into a stream of third party mentions you did not have to negotiate.
Make it quotable and make it checkable. Publish the sample, the dates, the method and the raw counts, not just the headline percentage. State the figure in one sentence that survives being lifted out of context. Keep the page crawlable to the agents that matter: OpenAI documents OAI-SearchBot and GPTBot as separate robots.txt controls with different purposes, and Anthropic documents ClaudeBot and the robots.txt directives that govern it, noting the rule has to be repeated per subdomain. Blocking everything is a decision, but make it deliberately. The page level craft is covered in our LLMO content playbook.
The earning motion: how to earn brand mentions in AI search, in order
Sequence matters more than effort here, because the cheap corrections raise the value of everything you do afterwards. Run it as five steps, in this order, and do not start step four until steps one to three are done.
One, fix the entity record. Organization markup, consistent naming, a clean about page, correct social profiles in sameAs. Two, fix every directory profile in your category, then start the review cadence. Three, audit partner and marketplace listings and send corrections. Those three are correction work, they take a fortnight, and they cost nothing but attention. Four, build community presence with named people on a standing weekly commitment. Five, pitch trade and analyst coverage and publish research, which are the long cycle plays that only pay off once the cheap layers are accurate. This is a sequence built to earn brand mentions in AI search that engines will actually reuse, rather than a content calendar dressed up as a strategy.
One warning about substitution. None of this replaces having something worth saying, and Google's guide to optimizing for generative AI features is explicit that creating content people find unique, compelling and useful will influence presence in generative AI search more than any other suggestion it makes. The five source types distribute a reputation. They do not create one. If you are working the recommendation angle specifically, how to get a SaaS product recommended by ChatGPT goes deeper on the shortlist mechanics.
How to tell whether it worked
Measure the outcome, not the activity. Freeze a set of thirty to sixty buyer prompts, never naming your own brand inside a shortlist prompt, run them on each engine on a schedule, and log one row per run plus one row per brand named. Run the identical set twice with no changes first, so you know the spread between two identical passes and can ignore movement smaller than it. The full build is in how to measure LLMO and AI visibility, and the same log tells you which AI citation sources keep appearing so you can attack the ones you do not yet appear on.
Two supporting numbers, used carefully. Google's generative AI performance report in Search Console shows impressions from generative AI features by page, country and device, and has been available to all websites worldwide since 31 August 2026, though a property with too few generative AI impressions may still see little data. Referral traffic is the weaker signal: Pew found users clicked a source inside an AI summary in just 1 percent of visits, and clicked any search result in 8 percent of visits with a summary against 15 percent without. Keep assistant referrals as their own caveated line using the method in tracking ChatGPT referral traffic in GA4, and never present a mention programme as a traffic programme.
Earn Brand Mentions in AI Search With DevCommX
DevCommX runs this as a GTM workstream rather than a content project: we audit every third party page that already describes you, sequence the corrections, build the prompt set and the run log with your sales team, and hand your marketers a target list with owners against it. We are a revenue operations and GTM engineering team first, which is the same operating discipline that produced 40+ qualified demos in ~6 weeks, from our AI SDR work on a fully scoped programme with a defined ICP. Book a working session and bring the last shortlist answer that left you out.
References
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, 22 July 2025, source for Wikipedia, YouTube and Reddit being the most commonly linked sources, for the 15 percent against 17 percent source share figures, for the 1 percent click rate on in summary sources and for the 8 against 15 percent result click figures. That analysis covers Google searches only, using browsing data from March 2025.
- Axios, ChatGPT Cuts Reddit Citations, source for Reddit falling from roughly 3.83 percent to 0.52 percent of ChatGPT citations between early and mid August 2026.
- Google Search Central, AI Features and Your Website, source for the rule that a page must be indexed and snippet eligible to appear as a supporting link in AI Overviews or AI Mode.
- Google Search Central, Organization Structured Data, source for sameAs being the URL of a page on another website with additional information about your organisation, and for iso6523 and naics being the properties used behind the scenes to disambiguate one organisation from another.
- Google Search Central, Creating Helpful, Reliable, People-First Content, source for the Who, How and Why self assessment questions on authorship and bylines.
- Google Search Central, Guide to Optimizing for Generative AI Features on Google Search, source for unique, compelling and useful content influencing presence in generative AI search more than any other suggestion.
- Google Search Console Help, Generative AI Performance Report, source for impressions from generative AI features by page, country and device.
- OpenAI, Overview of OpenAI Crawlers, source for OAI-SearchBot and GPTBot as separate robots.txt controls.
- Anthropic Help Center, Does Anthropic Crawl Data From the Web, and How Can Site Owners Block the Crawler?, source for ClaudeBot and its per subdomain robots.txt directives.
FAQ
How do I get my brand mentioned by AI?
Get described accurately on crawlable pages you do not own. In practice that means a complete directory profile, current partner and marketplace listings, genuine presence in the communities your buyers read, coverage from trade press or analysts, and original research that other people quote. Fix the pages that already describe you before commissioning anything new, because corrections are far cheaper than creation.
Do backlinks matter for AI search?
Links still help a page get crawled, indexed and trusted, so they have not stopped mattering. What has changed is that a mention with no link can still be lifted into an answer, because an assistant is reading the passage rather than following the anchor. Treat links and mentions as two outputs of the same outreach work instead of ranking one above the other.
Which sources do AI assistants cite most?
The best public evidence covers Google AI summaries rather than AI assistants generally, because Pew Research Center states that due to technical limitations its analysis includes only Google searches. In that browsing data from March 2025, Wikipedia, YouTube and Reddit were the most commonly linked sources in both Google AI summaries and standard search results, together accounting for 15 percent of the sources listed in the AI summaries it examined. That leaves most citations spread across a long tail, which is where category directories, trade titles and vendor documentation live.
Where do brand mentions in AI search actually come from?
Brand mentions in AI search come from ordinary crawlable web pages rather than from a special index. Five types carry most of them for B2B: review and comparison directories, practitioner communities, trade and analyst coverage, partner and integration pages, and original research that other publishers quote. Each is a page someone else controls, which is exactly why an assistant treats it as independent.
Do unlinked brand mentions count?
Yes, for this purpose they count. An assistant summarising a page can name your brand whether or not the page links to you, so an accurate unlinked description on a page buyers read is a real asset. Log unlinked mentions in the same tracker as links, record the exact wording used about you, and chase corrections when that wording is wrong.
How long does it take to earn brand mentions in AI search?
Directory and partner listing corrections can surface within weeks because those pages are recrawled often. Community presence and trade coverage run on human timelines of a quarter or more. Rather than forecasting a date, run a baseline pass of your prompt set now, repeat it on a fixed schedule, and report movement only when it exceeds the spread between two identical passes.
.webp)
