How to Get Your Brand Recommended by AI
The traffic charts are everywhere this month, and they all bend the same way. Clicks are draining out of the open web and into the answer box. Most founders are reading this as a traffic story and looking for the SEO tactic that wins it back. It is not a traffic story. It is a recommendation story, and the thing being handed out at the top of the funnel is no longer a ranked link. It is a name the model says out loud when a buyer asks what to use.
The problem: the click is dying and the recommendation is being born
Start with the numbers everyone is quoting, because they are real and they matter, then notice what they are actually telling you. About 68 percent of United States Google searches now end without a single click. When an AI Overview shows up at the top of the page, organic click-through on the links below it drops roughly 65 percent, and about a quarter of those users end the session right there, answer in hand. AI Overviews already trigger on close to 48 percent of tracked queries, up more than half year over year. And the chat assistants themselves, ChatGPT and the rest, send something like 95 percent less referral traffic to publishers than a normal Google result does, because their whole point is to answer inside the box instead of shipping you elsewhere.
Read as a traffic story, this is a disaster, and that is how most people are reading it. The reflex is to fight for the clicks that remain: tighten titles, add more pages, chase the shrinking pool of queries that still send a visitor. I understand the reflex. I have built companies where organic discovery was a real slice of growth, and I have watched those dashboards bend this year like everyone else’s.
Here is the reframe that changes what you do about it. The click was never the prize. The click was the delivery mechanism for the prize, and the prize was the recommendation, that moment where a stranger with intent gets pointed at you rather than a competitor. For twenty years the recommendation arrived as a ranked link, so we all agreed to call the link the goal and built an entire craft around earning it. The link is now optional. ChatGPT crossed 900 million weekly active users early this year. When one of them types “best tool for X” or “who should I hire to do Y,” the model does not hand back ten links for the person to sort. It names a short list of brands, usually two to seven, and most people take the names and move on.
So the real question is not how you claw back clicks. It is how you become one of the names. And the uncomfortable part, the part that breaks the old playbook, is that you cannot buy your way onto that short list the way you bought your way up the rankings. The lever is somewhere you are not used to pulling. This sits inside the larger map of where AI is moving value, and distribution is one of the squares moving fastest.
The framework: the recognition gap
Here is the single idea the rest of this rests on. A model recommends entities it recognizes, and it can only recognize an entity that shows up, described the same way, across enough independent sources that it can build a confident picture of who you are and what you are for. The distance between “the model has heard of you” and “the model is sure enough to name you” is the recognition gap. Almost every brand that is invisible in AI answers is not losing on quality. It is losing on recognition.
Think about how the answer is actually assembled. The model was trained on a large slice of the public web, so part of what it knows about you is baked in from that. Then, at the moment of the question, most assistants also retrieve fresh material and read it before answering. In both passes, the model is not grading your homepage against a rubric. It is asking a quieter question: across everything I have seen, who is this, and do enough sources agree that I can say their name without being wrong? If the answer is yes, you get named. If the answer is “I have seen the name a couple of times but I cannot pin down what they do or whether they are any good,” you get left out, and the model reaches for a brand it is more sure about.
That threshold is the mechanism, so it is worth naming plainly: the corroboration threshold is the point where enough independent sources say the same thing about you that the model treats it as fact rather than a claim. Below it, you are a maybe, and models are trained to skip maybes in favor of confident answers. Above it, you are a fact the model will repeat. Everything useful in this piece is a way to move more of what you want said about you from below that line to above it.
Notice how different this is from the SEO mental model, where the unit you optimized was a page and the goal was to make that one page win a query. Here the unit is the entity, meaning you, the company, the person, the product, as a thing the model can identify and describe. Pages still matter, but they matter as one source of corroboration among many, and usually not the most persuasive one, because the model knows you wrote your own page. I wrote a broader piece on generative engine optimization for founders that maps the whole surface. This one goes narrow and deep on the part almost everyone gets wrong: recognition is built off your site, not on it.
The three signals a model reads
When researchers and practitioners take apart what actually moves a brand into AI answers, the same three signals keep surfacing. I think of them as identity, corroboration, and consistency, and they stack in that order.
Identity is whether the model can tell what you are at all. Not your tagline, your category and your specifics: what you do, who you serve, what makes you different, stated in plain declarative sentences that a machine can lift without guessing. Corroboration is whether other sources, ones you do not own, say the same thing. This is the heavy one, and it is where most of the work lives. Consistency is whether your name, description, and identifiers line up everywhere they appear, so the model does not split you into two half-formed entities or confuse you with a similarly named company. A model that cannot decide whether “Acme” is your analytics startup or a coyote’s hardware supplier will quietly route around the ambiguity and recommend something cleaner.
| Signal | What the model is checking | How you build it |
|---|---|---|
| Identity | Can I tell what this is, who it serves, and what makes it distinct? | Plain declarative facts about your category and edge, stated the same way everywhere. Organization schema and a clear about page as the seed. |
| Corroboration | Do sources I do not think you control say the same thing? | Earned mentions on third-party sites, reviews, forums, video, and press. The more independent, the heavier the vote. |
| Consistency | Is this one stable entity, or a smear of conflicting descriptions? | Same name, same one-line description, same identifiers and sameAs links across every profile, listing, and byline. |
The order is not decorative. Consistency without identity is a well-organized blank. Identity without corroboration is a confident claim the model has no reason to trust. You need all three, and the one that is scarce, the one you cannot manufacture on your own domain in an afternoon, is corroboration. Which brings us to the flip that reorganizes the whole job.
The off-page flip: the lever moved off your website
For two decades, the most powerful surface in discovery was the one you owned. You controlled your pages, your structure, your internal links, and you earned backlinks pointing in. The craft was real and the levers were mostly yours to pull. AI answers invert that. The signals that best predict whether a model cites or recommends you are the ones you do not control, and the signal you spent years accumulating, backlinks, has slid down the list. This is the same pattern I traced in the commoditization clock: the thing that used to be your edge becomes table stakes, and the edge moves to whatever is harder to manufacture. Backlinks got manufacturable. Genuine third-party consensus did not.
The clearest evidence comes from a study of roughly 75,000 brands that correlated various signals with AI citations. Off-site brand signals dominated. Mentions on video correlated about 0.737 with getting cited, plain branded web mentions about 0.664, branded anchor text about 0.527, and branded search volume, meaning how many people look you up by name, about 0.392. Traditional backlinks came in around 0.218, roughly two to three times weaker than the brand signals sitting above them. Separately, analyses of AI Overview citations found that earned media accounts for the large majority of what gets cited, on the order of 82 to 89 percent, and that only about 38 percent of the pages an AI Overview cites even rank in Google’s own top ten for the query. The old proxy and the new one have come apart.
Sit with what that means for how you spend the next quarter. If your entire plan is on-page, better titles, more schema, a tidier site, you are working the weakest lever on the board and calling it strategy. The heavy levers are a mention in a video someone else made, a thread on a forum you do not run, a review on a site you cannot edit, a name-drop in an article you did not write. This is closer to public relations and community than to classic search work, and that is not a metaphor, it is the actual shape of the job now.
Why your own website stopped being the lever
It helps to understand the plumbing, because it explains why the flip is structural and not a passing quirk of one model. Language models learn who you are from two places, and neither of them is impressed by your homepage on its own.
The first place is training data, the giant corpus scraped from the public web and structured sources. A large share of what a model treats as settled fact about the world comes through a small set of high-trust references, and two of them do a lot of quiet work: Wikipedia and Wikidata. Wikidata is the machine-readable layer that feeds Google’s Knowledge Graph and a chain of assistants behind it, and it is part of the training diet for most major models. If those references have never heard of you, the model’s baseline picture of you is thin, and it fills the gap with whatever it can corroborate elsewhere. This is why a brand can rank number one on Google for its own category and still be a ghost inside ChatGPT. Ranking is about your pages. Recognition is about whether the wider record knows you exist and agrees on what you are.
The second place is retrieval at question time, where the assistant pulls fresh sources and reads them before answering. Here your site can appear, but it appears as one voice among many, and the model discounts self-description in the presence of third-party corroboration. It has learned, correctly, that every company says it is the best. What moves the needle is when sources the model does not think you control say something specific and consistent about you. Your website is table stakes now: it needs to exist, be clear, carry clean Organization schema, and state your facts plainly so they are easy to lift. But the site is the floor, not the lever. Treating schema markup as the growth plan is the same category error as the wrapper trap, where a thin layer over someone else’s engine gets mistaken for a business. Clean schema is hygiene. Corroboration is the work.
The mention economy: how recognition actually gets built
If corroboration is the currency, the practical question is how you mint it, because you cannot post it on your own domain and cannot buy it in bulk without the model noticing the pattern. Recognition gets built the slow way, through mentions that other people choose to make, and the founders who win at this treat earned mentions as a first-class growth output rather than a byproduct of vanity.
Mentions come from being worth mentioning, which sounds soft until you break it into moves. The reliable sources of durable mentions are a short list. Be quotable: publish a real point of view, a named framework, a specific number nobody else has, so that when someone writes about the topic you are the convenient thing to cite. Show up in the places buyers already ask each other for recommendations, which for most categories means forums like Reddit, communities, and review sites like G2 where a genuine track record accumulates. Get on other people’s surfaces: a podcast, a guest piece, a conference talk, a collaboration, anything that plants your name and description on a domain that is not yours. Publish original data, because a statistic with your name attached gets repeated, and every repeat is a corroborating vote. And earn the occasional piece of real press, because news and trade coverage carry disproportionate weight in what a model treats as established.
There is a compounding property here worth planning around, because it changes how you pace the work. The first few mentions do almost nothing on their own, since the model has no reason to trust one stray source. The value shows up nonlinearly, when the count of independent, consistent mentions crosses the point where the model stops treating your claim as a claim and starts treating it as fact. That is why founders who dabble give up right before it would have worked, and why the ones who commit to a quarter of steady corroboration suddenly appear in answers that ignored them for months. It is a threshold effect, not a dial, and thresholds reward patience that dabbling never earns. The economics also favor it: once a model reliably names you, that recommendation costs nothing per query and does not deflate the way paid channels do, which makes it one of the few distribution assets that gets cheaper to hold even as the price of the underlying inference keeps falling.
The thread running through all of that is that you are building distribution you do not own, on purpose, which feels backward to a generation of founders taught to own the channel. It rhymes with how agents are becoming a new front door, where the buyer never touches your site because an assistant handled the whole errand. When the buyer is a model, or a person taking a model’s word, the mention on someone else’s turf is the distribution. Owning your channel still matters for the customers you already have. It does almost nothing to make a model recommend you to the ones you do not.
| Dimension | SEO (winning a link) | Getting recommended by AI (winning a name) |
|---|---|---|
| Unit of visibility | The page | The entity (you, the company, the product) |
| Where the lever sits | On your site, mostly under your control | Off your site, mostly out of your control |
| Currency | Backlinks and rankings | Independent mentions and corroboration |
| What you optimize | Content and structure | Reputation and consistency of description |
| The metric | Rank position and organic traffic | Share of model: how often you get named |
| Feedback loop | Crawl and index, days to weeks | Training and retrieval, weeks to months |
Two founders and one category question
Make it concrete, because the abstraction hides how differently this plays out for two companies that look identical on paper. Picture two founders in the same category, both with a good product, both ranking respectably on Google, both with clean, well-structured sites. A buyer opens ChatGPT and asks the plain question: what are the best tools for this job.
The first founder spent the year on-page. More landing pages, more schema, more FAQ blocks written for the model, a keyword refresh every month. Real work, all of it on a surface she controls. The model reads her tidy site during retrieval, but when it assembles the answer, it reaches for the names it is sure about, and hers is not one of them, because nothing outside her own domain vouches for what she does. She is clear and uncorroborated, the known-but-thin corner, and she never gets said out loud.
The second founder spent the year off-page. He published a named framework and an original benchmark, got himself on three podcasts his buyers listen to, seeded honest presence on the review sites and the one forum where his category argues, and made sure his one-line description was identical everywhere. His own site is merely fine. But when the model gets the category question, six independent sources have already told it, in roughly the same words, what he does and that he is good at it. He crosses the threshold. The model names him, describes him correctly, and hands him a buyer who is already three-quarters sold. Same product, same rankings, opposite outcome, and the only difference was where they aimed the year. This is also why the direct relationship still matters once you win it. A recommendation gets you the introduction, but the moment the buyer becomes yours, you want the switching costs and the relationship to be real, the same logic behind designing for the agent as your first customer and thinking hard about where switching costs actually live.
The recommendation test and share of model
You cannot manage what you refuse to measure, and the number most people reach for, AI referral traffic, is close to useless as a steering signal because it is small and noisy and does not tell you whether you got named. So replace it with two better instruments.
The first is the recommendation test, and it takes five minutes. Open the assistants your buyers actually use, ChatGPT, Google’s AI mode, Perplexity, and ask the questions a buyer would ask at the moment of choosing: “what are the best tools for X,” “who should I use to do Y,” “compare the top options for Z.” Do not ask about yourself by name, that is cheating, the model will find your site and read it back to you. Ask the unprompted category question the way a stranger would. Then read the answer for one thing: did it say your name, and if so, was the description right? That binary, named or not named, is the truest read of where you stand, and it is brutal the first time you run it.
The second instrument is share of model, the percentage of a fixed set of buyer questions where you get named, tracked over time. Build a list of twenty or thirty real questions your buyers ask, run them across the main assistants on a schedule, and count how often your name appears and how often each competitor’s does. That single ratio is your rank in the new world. It is the answer-engine equivalent of share of voice, and it is now sold as a category of tooling with pricing tiers and enterprise contracts, because everyone realized at once that positional rankings stopped describing reality. You do not need to buy a tool to start. You need a spreadsheet, thirty questions, and the discipline to run them monthly. The map above tells you what a low score means: if the model cannot even describe you, your problem is identity, and no amount of corroboration lands until you fix it; if it describes you fine but never names you, your problem is that too few independent sources vouch for you, and that is a corroboration campaign, not a copy edit.
The one percent that is not what it looks like
Now the honest counterweight, because there is a real argument on the other side and it deserves a fair hearing. Add up the clicks that assistants send back to sites and it is tiny. Across a broad sample of industries, AI referral traffic is on the order of 1 percent of total visits. A rational founder looks at 1 percent and asks why any of this deserves a quarter of attention. If you were measuring in raw traffic, they would be right to shrug.
But raw traffic is the wrong ruler, and the conversion data is why. AI-referred visitors arrive further along in their decision than a random searcher, because the model already did the shortlisting and handed them a recommendation. The result shows up in the numbers with unusual consistency. Depending on the study, AI-referred visitors convert somewhere from four to more than twenty times better than standard organic traffic. One clean read put sign-up conversion at 1.66 percent for AI-visible traffic against 0.15 percent for ordinary organic, roughly an order of magnitude. Case studies keep landing in the same place: a B2B software team saw AI-referred leads convert at close to three times their organic rate, and conversion lift compounds further when a brand is named across two or more assistants at once. These are self-reported and they vary, so hold them loosely, but the direction is stable across every dataset I have seen.
The point is not that AI referral traffic is secretly huge. It is that traffic was never what you were buying. You were buying the recommendation, and the recommendation now arrives pre-qualified, which is exactly why the click count looks small and the conversion looks large. Measuring this channel in visits is like measuring a great sales referral by how many minutes the introduction email took to read. This is the same posture I keep coming back to on platform risk: the number that looks reassuring is often the wrong number, and the one that matters is quieter and harder to game.
The contrarian take: GEO is a PR problem in an SEO costume
Here is where most of the advice goes wrong, including a lot of it dressed up as the new thing. The market took the arrival of AI search and handed it to the SEO team, because it looks adjacent and the words rhyme, and the SEO team did the reasonable thing and ran their playbook harder. More schema. More FAQ blocks. More keyword-shaped pages, now written for the model instead of the crawler. It feels like progress because it is motion on a surface you control. It mostly does not move the number, because it is all on-page, and on-page is the weak lever.
The uncomfortable truth is that getting recommended by AI is a public relations and reputation problem wearing an SEO costume. The engine of it is other people saying specific, consistent things about you on surfaces you do not own. That is the job of communications, community, category leadership, and a real point of view, far more than the job of technical optimization. The teams that will win the next few years are the ones that reorganize around that fact instead of fighting it, the way the sharp ones already reorganized around taste and judgment when execution got cheap.
And to argue the other side honestly, because the SEO-is-the-same crowd is not entirely wrong: the fundamentals do transfer. Clear structure, extractable answers, clean markup, and pages that state facts plainly all genuinely help a model lift and trust your content, and a brand with a broken, unreadable site will underperform its reputation. So do the on-page hygiene. It is real and it is cheap. The mistake is stopping there, treating the hygiene as the strategy, and never funding the off-page work that actually crosses the corroboration threshold. Hygiene gets you eligible. Corroboration gets you named.
What to do Monday morning: the entity-building audit
Turn all of this into five moves you can start this week. None of them require a tool you have to buy, and the first one takes five minutes.
First, run the recommendation test. Take your ten most important buyer questions, ask them unprompted across ChatGPT, Perplexity, and Google’s AI mode, and write down whether you were named and whether the description was right. That is your baseline, and it will probably sting. Second, fix identity and consistency, because it is the cheapest lever and it gates the rest. Write one clean sentence that says what you are, who you serve, and what makes you different, and make it identical across your site, your Organization schema, your profiles, your directory listings, and every place your name appears. Add sameAs links so the machine record ties your presences together into one entity. Third, launch a corroboration campaign with a target you can count: get named and correctly described on at least three independent third-party surfaces this quarter, a review site, a forum thread, a guest piece, a podcast, a comparison article, anything you do not own. Pick the ones your buyers already read. Fourth, earn one durable mention that carries weight, an original data point with your name on it, a piece of real press, or a strong guest appearance, because one heavy corroborating source can outvote a dozen light ones. Fifth, stand up share of model as a tracked metric: thirty buyer questions, run monthly, count how often you and each rival get named, and set a tripwire with a date, for example, “if we are not in the model’s short list for our top five questions by the end of next quarter, corroboration becomes a company priority, not a side project.”
Sequence matters as much as the moves. Identity and consistency come first because they are cheap and because corroboration lands on nothing if the model cannot tell what you are. Corroboration comes next because it is the scarce currency and the slow one. Measurement wraps around all of it so you can see whether the work is landing before you have spent a quarter guessing. Do those five and you are working the levers that actually move recognition, in the order that makes them compound. Skip them and keep polishing pages, and you will have the tidiest site nobody’s assistant ever mentions. When discovery gets re-platformed, the winners are not the ones who optimized the old surface hardest. They are the ones who noticed the prize moved and went to where it landed. The prize moved off your website. Go build where the model is actually looking.
Frequently asked questions
What is the difference between SEO and getting recommended by AI?
SEO optimizes a page to win a ranked link for a query, using levers mostly on your own site, and measures success in rank position and traffic. Getting recommended by AI optimizes an entity, meaning you or your company as a thing the model can identify, so that the model names you in its answer. The main levers sit off your site, in independent mentions and corroboration, and success is measured by how often you get named, not how much traffic you get.
Why does my brand rank on Google but never show up in ChatGPT?
Ranking is about your pages. Recognition is about whether the wider public record knows you exist and agrees on what you are. A model builds its picture of you from training data, where high-trust references like Wikipedia and Wikidata carry weight, and from third-party sources it retrieves at answer time. If those sources are thin or inconsistent about you, the model cannot describe you confidently and skips you in favor of a brand it is more sure of, even if you outrank that brand on Google.
Do I need a Wikipedia or Wikidata page to get recommended by AI?
It helps but it is not required, and you should not fabricate one, because Wikipedia has notability rules and manufactured entries get removed. Wikidata and Wikipedia feed the knowledge graphs and training data behind many assistants, so a legitimate presence strengthens your baseline recognition. If you do not qualify yet, the substitute is breadth of independent corroboration across reviews, forums, video, press, and directories, all describing you consistently.
How do AI models decide which brands to recommend?
They favor entities they can recognize with confidence. That confidence comes from three stacked signals: identity, whether the model can tell what you are and who you serve; corroboration, whether independent sources it does not think you control say the same thing; and consistency, whether your name, description, and identifiers line up everywhere. When enough independent sources agree, the model treats it as fact and repeats it. Below that corroboration threshold, it hedges and picks a safer name.
Is AI search worth it if it is only about 1 percent of my traffic?
Measured in raw traffic, it looks small, around 1 percent of visits across many industries. Measured in outcomes, it is not, because AI-referred visitors arrive pre-qualified by the model’s shortlisting and convert several times better than ordinary organic, in many studies four times or more, and by one read roughly ten times on sign-ups. You are not buying traffic, you are buying the recommendation, which is why the click count is small and the conversion is high.
What is AI share of voice and how do I measure it?
AI share of voice, or share of model, is the percentage of a fixed set of buyer questions where the model names your brand, compared with competitors, tracked over time. To measure it, build a list of twenty to thirty real questions your buyers ask, run them monthly across the main assistants, and count how often you and each rival appear. It is the answer-engine replacement for rank position, and you can start with a spreadsheet before buying any dedicated tool.
Can I just add schema markup and FAQ blocks to get cited?
Those help, but they are hygiene, not strategy. Clean Organization schema, clear structure, and extractable answers make you eligible to be lifted and trusted, and a broken site will underperform. But on-page work is the weakest lever for AI recommendation, because models discount self-description and weigh independent corroboration far more heavily. Do the hygiene, then fund the off-page work, earned mentions and consistent third-party description, that actually crosses the threshold.
How long does it take to get recommended by AI search?
Longer than an SEO change and shorter than building a brand from nothing. Identity and consistency fixes can register within weeks as assistants retrieve fresh sources. Corroboration is slower, because it depends on other people publishing mentions and, for the training-data layer, on future model updates ingesting them, which runs on a cycle of weeks to months. Treat it as a compounding campaign measured in quarters, not a one-time optimization, and track share of model so you can see it move.