Where AI Looks Before It Recommends You
Picture two plumbers in the same town. Same work, same prices, same twelve years on the tools. One has a 4.2 rating, the other has a 5.0. On an average Tuesday, the 4.2 gets more calls.
Five years ago you would have skimmed Google's list, picked the highest rating and rung whoever answered. That is not what happens now. People increasingly ask ChatGPT or Gemini to recommend someone, and those systems do not choose the way you would. Review sites are one of the first places they look, but a lot of review content sits behind a paywall they cannot read. So they work from what they can see: how many reviews there are, the star rating, and whatever text is not blocked. The plumber with 300 visible reviews at 4.2 looks safer than the one with nine at 5.0, and the recommendation follows.
That is the whole point of this piece. AI does not weigh sources the way a person does, so it is worth knowing which ones it actually leans on before you decide where to spend your effort.
Our ranked view of what AI leans on
Ordered from most weight to least, in our view.
| Source | Weight | Why we rate it there |
|---|---|---|
| Wikipedia | Top tier | Treated as a source of truth. If it is stated there, the model tends to accept it and build on it. |
| Media coverage and digital PR | Top tier | External sources talking about you compound authority strongly. One mention helps. A pattern of them across independent outlets is what really moves things. |
| Review sites (Trustpilot, Google reviews) | High | Third-party verification of quality, which is exactly what a model is trying to find when someone asks it who to trust. |
| Industry publications | High | Niche relevance makes them credible. A trade title in your field carries more weight on your subject than a bigger, broader outlet. |
| YouTube | Upper middle | Substantial and well indexed, and it puts a face and a demonstration behind the claim. |
| Podcasts | Middle | Good, but it depends heavily on the size of the show. A small podcast is barely a signal. |
| Awards | Middle | Real, but only one signal among many. An award on its own does not carry a recommendation. |
| Directories (Clutch and similar) | Middle | Used, but not weighted heavily. Worth having, not worth building a strategy around. |
| Your own website | Middle | Lower than most people expect. It is the foundation, not the proof. |
| Middle, and falling | It used to rank far higher in our view and is being pulled down. We would not build a plan on it holding its position. | |
| Low | Anyone can publish their own claims there, so it carries little weight. Self-published is self-published. |
The shape of that list matters more than any individual row. The things sitting at the top are all things other people say about you. The things at the bottom are mostly things you say about yourself.
The half that actually counts
Here is the conclusion, and it is uncomfortable if you have spent two years writing pages. Everything on your own site is you saying you are good. That is not nothing. It is the foundation, and without it there is nothing for anyone to check. But it is only ever half the picture, and it is the half a model discounts, because every one of your competitors is making the same claims on their own sites.
Third parties saying it is the half that counts. A journalist, a reviewer, a trade publication, a customer leaving a review under their own name: none of those are under your control, which is precisely why they are worth something.
That reframes the work. The instinct when rankings stall is to write more pages. In our view the better question is: where am I mentioned that is not my own website, and how do I get mentioned in more of those places? We have written about the wider shift in SEO in the age of AI, how citations get chosen in AI Overviews SEO, and where the market is heading in the state of AI search in 2026.
What to do about it
Make sure AI can actually read your site
This comes first because authority is worthless if you are invisible. If your robots rules block AI crawlers, or your content only appears after JavaScript runs, or your pages return errors to anything that is not a browser, none of the rest of this matters. Check what you are allowing with the robots directives checker and the background in what is robots.txt, then run a site audit to confirm your pages are reachable and returning what you think they are. Clean structured data helps too: it tells a machine what your page is about without making it guess.
Publish something only you have
Your own numbers, your own results, your own experience. This is the one thing a model cannot get anywhere else, and when it needs that information it has to point at you to give it. Aggregated advice that exists on forty other sites gives it no reason to name you. A figure from your own book, a result you measured, a method you worked out the hard way: that is citable. Our guide to writing SEO content covers how to build a page around it, and E-E-A-T explains why first-hand experience is treated as a quality signal in the first place.
Keep your details consistent everywhere
Same business name, same address, same phone number, on every listing, profile and directory entry. AI cross-checks these, and listings that disagree with each other make you look unreliable, or worse, make you look like two different businesses. This is dull, unglamorous work and it is one of the few things on this list you can finish in an afternoon. The local SEO guide covers how to work through it.
Get other people talking about you
This is the one that moves the needle, and it is the one everybody avoids because it is slow and it involves other people. Pitch the trade press. Give a journalist a quote worth using. Ask satisfied customers to leave a review where it is visible rather than behind a paywall. Go on the podcast. Earn the mention properly rather than buying it. Start with how to build backlinks and what link building is, and if you publish video, the YouTube SEO guide is the relevant one.
None of this is a trick. It is the same thing reputation has always been, measured by a machine that reads faster than you do and trusts you less than you would like.
Last updated: September 2026