Being crawled, being cited, and being chosen are three different things. Most organizations track one of them and assume it covers the other two.
Clients keep asking a version of the same question: how do we show up in AI? For a while we could only answer in generalities. A site can rank perfectly well and still lose ground in AI, and the analytics most teams already run will not show it, because those tools were built for a web where people click links, not one where a machine reads the page and answers for them. In the first four months of 2026, 68% of US Google searches ended without a click to any website (SparkToro), a shift we started tracking when AI Overviews first began reshaping search traffic.” The answer a person gets is assembled mostly from websites, and your own site is the one source in that mix you control.
So we ran the audit. We looked at how AI sees one organization’s website, a credit union, across five AI assistants.
Being read is not being recommended
Crawling means an AI system downloaded your page. Citation means it named your site as a source in an answer. Selection means it recommended you over an alternative. Each depends on different signals, each is measured differently, and strength in one says nothing about the other two.
AI assistants were reading the site constantly. Microsoft’s crawlers accounted for most of that activity, with OpenAI, Anthropic, Google, Meta, and Amazon all present at lower volume. That looks like progress, and on its own it isn’t. Heavy crawling only means the machines are collecting your content. It says nothing about whether they recommend you.
The gap between the two is wide. Cloudflare, which sits in front of a large share of the web, tracks how often each AI platform crawls a site against how often it sends a visitor back. The pattern holds across platforms: AI crawlers request orders of magnitude more pages than they return visits (Cloudflare). Traditional search sits far closer to an even exchange. Treat heavy AI crawling as proof that AI is recommending you, and you will be confident and wrong.
Present everywhere, chosen nowhere
Here is the finding that should get a marketing leader’s attention. On the prompt set we tested, this brand was cited more often than any other domain in its category. Read quickly, that is a win.
The brand had broad presence, showing up and often ranking first on general, informational questions. But its citations carried real weight only where the person already knew the brand and was looking for it: its name, its website, its branches. On the questions that actually win new business, the “which credit union should I choose” and “best credit union near me” comparisons, its authority fell by about half. Our own prompt panel put two competitors ahead of it on those questions, and on two of the five assistants it dropped out of about one answer in five entirely.
So the brand was everywhere someone already knew it, and nowhere a new prospect was deciding.
That gap is not a technical problem. On comparison questions, AI leans on third-party sources far more than on your own site: community discussion like Reddit, video like YouTube, which now draws more AI citations than Reddit does (eMarketer), and review platforms like Google reviews and Yelp (Search Engine Land), along with “best of” listicles and comparison pages, which take about 40% of citations on commercial-intent questions, close to double any other format (Search Engine Land). Winning those is brand and reputation work, your job, guided by your own data. The technical layer’s role is narrower but essential: making sure that when those signals exist, the machines can read them and attribute them to you.
Whether AI describes you accurately is a brand-authority question
One fix costs almost nothing, and almost none of the competitors we checked had made it: giving machines a clear, structured account of who you are. That means Organization structured data, the markup that states your name, services, and scale in a form a machine reads directly. Call it your machine-readable identity. Without it, AI describes you using whatever third parties say, because you never gave it an authoritative version to prefer.
Be clear on what this does and does not do. It will not, on its own, get you cited more often. Ahrefs tracked 1,885 pages that added structured data against about 4,000 matched pages that did not, measuring citations across Google AI Overviews, Google AI Mode, and ChatGPT. On AI Mode and ChatGPT the lift was statistically indistinguishable from zero, and on AI Overviews the pages that added schema slipped slightly behind the controls (Ahrefs). Anyone selling structured data as a citation hack is overselling it.
The same researchers did find that pages AI cites are almost three times more likely to carry the markup. They attribute that to the kind of site that invests in markup, not to the markup itself. Correlation, not cause.
What structured data does is control accuracy: whether an assistant gets your name, your services, and your scale right, or repeats something stale and wrong. For an organization that has changed what it does or how it is positioned, that accuracy is the point. This is brand authority in a basic sense, whether you or a third party gets to define who you are to the machine. It is cheap, it is largely ignored by competitors, and that combination makes it a rare advantage.
AI remembers the old you
Models keep repeating facts that are no longer true. On the site we audited, AI answers still described an earlier version of the organization, smaller and narrower than what it is now. We also found AI citing pages that had been retired well before the audit. You cannot edit a model’s memory, the same way you cannot edit a person’s. What you can do is publish a clear, current account of who you are now, so the next time AI reads you, it has something better to learn from. Without that current source of truth on your own site, nothing authoritative overrides the outdated version.
Two lanes, and where to start
The work splits cleanly into two lanes. The technical lane is your site’s structure, its machine-readable identity, and honest measurement. That is an agency’s job, and it is where COLAB works, including the work we do with credit unions. The other lane is comparison content, reviews, and community presence, the signals that win the high-intent decision. That is your brand’s job, guided by your own data. The common mistake is confusing the two, either buying a tool and calling it a strategy, or pouring effort into content while the machines still cannot tell who you are.
Start where the cost is lowest and the return is clearest. Give the machines a correct, structured account of who you are. Then check again in a quarter, and look for whether AI describes you accurately rather than whether it cites you more often. Accuracy is the thing you can actually move.
One question your team will raise: how do we measure any of this? The data exists, spread across your server logs, the AI answers themselves, and prompt-sampling tools. Assembling it into one honest picture is the hard part. That answer is technical, and we wrote it up separately for the people who own it. If your IT or web lead needs the how-to, send them our companion guide, How to Measure AI Visibility.
If your rankings look healthy but your AI presence is a question mark, that gap is worth understanding before it widens. Our AI services team runs this analysis on your actual site. If “how do we show up in AI?” is already on your desk, let’s talk about the answer for your organization.
