How AI Decides What To Recommend: A Starter's Guide

At some point in the last year you typed your own company name into ChatGPT to see what it would say. Everybody has. And depending on what came back, you either closed the tab feeling fine, or you sat there for a second wondering how a machine that has read most of the internet managed to describe you as something you stopped being three years ago, or skipped you entirely and recommended a competitor you have beaten on every dimension that matters.

That gap between what you are and what the machine thinks you are is the whole subject here. It is not a branding problem and it is not a content problem exactly. It is a records problem, and records problems are fixable.

I gave a version of this talk twice this summer, once at the AI Cafe at The KINN in Venice and once at Golden Hour, and this is the whole thing written down so you can work through it without me. I run content and authority strategy at Product.ai (formerly Demand.io), and everything in here is work I have actually run, mostly by getting it wrong first. One promise up front, because it sets the tone for the rest: nobody is a ten-year expert in a two-year-old field, including me, which is exactly why every method here is one you can check yourself, for free, this week.

What actually changed about getting found?

For about twenty years, getting found meant convincing Google that your page was the best answer to a query. Now an assistant answers first, and it builds that answer mostly out of what other people have published about you rather than what you have published about yourself.

The version I say on stage:

SEO is what you say about yourself. AEO is what everyone else says about you — and the machine trusts the second one a lot more.

AEO is answer engine optimization, the practice of getting picked up by AI assistants rather than ranked by a search engine. I did not invent that framing, by the way. It is the most useful way I have heard it put and I picked it up from someone else.

The second thing that changed is the surface area. It used to be search engine optimization, one box and one engine. Now the machine reads your reviews, your Reddit threads, your community, your support pages, your About page. Somebody in Neil Patel’s orbit started calling it search everywhere optimization, which is the honest name for it. I do not know who coined it first, but that is where I found it.

Practically, that means everywhere is now your responsibility. You cannot neglect your review profiles. You cannot neglect the forum thread where somebody asked whether you are any good. The model is already assembling a story about you from all of it, whether or not you participate in writing it.

What is visibility engineering?

Since I keep needing a name for the work itself, here is the one I have started using. Visibility engineering is the practice of making good work findable on purpose: building the record, the corroboration, and the page structure a machine needs before it will recommend you, rather than publishing into the void and hoping the right person stumbles across it.

This is not a campaign you run and it is not a message you broadcast. It is a system with parts you can inspect, and the parts fail in specific, diagnosable ways. That is good news, because it means the answer to “why doesn’t AI know us” is almost never “we need to be louder.”

One rule sits underneath all of it, and it is the reason I do this work rather than the other kind: visibility without value is just manipulation. The order matters. Be worth finding, then be findable. Reverse those and you are building a machine for tricking people, and the machines that read everywhere are getting steadily better at smelling it.

How does Google decide, and how does an AI decide?

These are two different technologies and the difference explains almost everything downstream.

Google sends out an army of robots that crawl the web by following links from page to page, then ranks the pages it finds, partly by how many other pages link to them. A link is basically a vote. That link-voting idea is PageRank, and it is the thing that built the company. The important part for you: Google’s unit is the whole page. It ranks pages.

An AI model is not doing that. It is not a database and it does not look things up in the way you are picturing. It read an enormous chunk of the internet and now it predicts the most likely next words. It is the world’s most well-read autocomplete. That is also why it sometimes invents things: a confident wrong answer gets assembled exactly the same way a confident right one does, because the model is predicting what sounds correct rather than checking a record.

Which leads to the consequence that earns the title. These two machines keep two different lists. Ahrefs ran 15,000 prompts through Google and the major assistants and found that, on average, only about 12% of the URLs the AIs cited ranked in Google’s top 10 for the same prompt, and roughly 80% did not rank in the top 100 at all. You can be page one on Google and invisible inside ChatGPT on the same day.

Which does not mean SEO is dead, and anyone telling you it is has something to sell. iPullRank looked at 79,000+ URL-query pairs and found ranking position still functions as, in their words, the “great gatekeeper” of AI citations, with a stark drop-off outside the top 10. Ranking gets you in the room. It does not decide who gets picked.

Why does AI quote one paragraph instead of ranking your page?

Before an assistant can use your page, the page gets chopped into chunks, and only the chunks get stored and searched. Not the whole document. It is a library that tears every book into single pages and files the pages instead of the books. If your one good answer is spread across three paragraphs, the machine may grab the middle one and miss the point entirely.

So Google grades the whole essay and the AI rips out the one paragraph it needs and quotes it. Make sure you have paragraphs worth ripping out.

What a liftable paragraph looks like, concretely: a real question as the heading, a tight answer of roughly forty to sixty words directly beneath it, facts first, no wind-up. In Kevin Indig’s analysis of 1.2 million ChatGPT answers covering 18,012 verified citations, cited content was twice as likely to contain a question mark, 78.4% of those question-tied citations came from headings, and 44.2% of all citations came from the first 30% of the page.

The machine reads the top. Most of us bury the answer under three paragraphs of throat-clearing because that is how we were taught to write, and every one of those paragraphs is pushing the good part out of range.

You do not need to rewrite your website. Take the two or three pages tied to how people actually buy from you, and pull the real answer up to the top of each one.

What are the two gates of AI visibility?

This is the model I would keep if you threw out everything else on the page.

Gate one is entity. Does the machine know you exist as a real, distinct thing, separate from everything else that shares your name?

Gate two is content. Once it knows who you are, can it lift a clean answer off your pages?

They multiply. Fail either one and you get zero. Beautiful, perfectly structured content attached to an entity the machine cannot resolve produces nothing at all, and a rock-solid entity with unliftable pages produces silence.

Press without entity is noise. Entity without press is silence. You need both, in that order.

Most people are working hard on gate two while still locked out of gate one, and a great deal of gate-two work gets sold to people who never passed gate one. If you take one sequencing rule from this piece: fix gate one before you spend a dollar on gate two.

Here is what failing gate one actually looks like, from my own year. One of our brands is SimplyCodes, a coupon platform. There is a coding school out there with a nearly identical name. To a machine those two names sit right next to each other, so when somebody asks an assistant whether SimplyCodes is legit, the model has to work out in a split second which thing you even mean. When the signals are muddy it does the safe thing: it gets less confident about both of us and recommends a competitor it feels sure about instead. That cost me a chunk of January. The fix was not marketing. It was going into the databases machines treat as ground truth and drawing a hard line saying this entity is not that one.

You are not writing marketing copy anymore. You are teaching a machine who you are so it stops confusing you with someone else.

Who does the machine actually trust?

Start at the bottom of the ladder, because the bottom is you. The least trusted source about your company is your company. The models are looking for neutral parties, on the reasonable theory that you will say anything about yourself.

Above you, in rising order of trust: a handful of boring old databases the models treat as ground truth, then review platforms, then earned press, then forums and communities where real people talk without being paid to.

The size of that gap surprised me. Muck Rack analyzed more than 25 million AI citations across ChatGPT, Claude, and Gemini in 17 industries and found 84% went to earned, third-party media. Paid and advertorial content accounted for 0.3%. Researchers at the University of Toronto found the same structural bias, with brand-owned pages under roughly 20% of citations in category after category.

Your own website is one of the least-cited sources about you.

Wikipedia sits unusually high on that ladder. It is ChatGPT’s single biggest source, nearly half the citations among its top-cited domains and about 13% of all U.S. ChatGPT citations in 2026 data. You cannot write your own page, and trying gets it deleted. But its structured sibling, Wikidata, is a different story, and I will get to it.

What should you do Monday?

Three things, and they go in this order, because each one makes the same single thing unmistakable: your entity.

Before any of them, run the free audit so you are working from evidence instead of a hunch. That is its own piece, and it is genuinely step zero.

1. Claim it. Get into the databases machines treat as ground truth. Wikidata first: it is the structured-facts database behind Wikipedia, the bar is far lower than Wikipedia proper, it is largely in your control, and creating a clean entry takes about half an hour. It also has a property that literally means “different from,” which you point at your look-alike to tell the machine you are not them. Then find the five or ten boring old databases your specific field runs on. Companies have Crunchbase and G2 and Trustpilot. Writers have Open Library and Goodreads and Amazon Author Central. Artists have Artsy and Behance and gallery pages. Musicians have Discogs and MusicBrainz. Local businesses have Google Business Profile and Yelp. If you do not know yours, ask an assistant what the authority files are for your field and check the answer across two or three models.

2. Confirm it. Make your own site tell the same story. Your homepage, About page, and FAQ should carry the same name, the same one-line description, and the same links as everything you just claimed. Write those pages as questions with plain answers under them, including the unflattering ones. Work backwards from the machine’s actual misreads: ours had decided SimplyCodes was a cashback company, which it is not, so the FAQ now answers that flat. Answer the questions before people have to ask them.

3. Earn it. Get a few genuinely trusted third parties in your specific niche to say something about you. Not the giants. The small, trusted ones your customers actually read: niche blogs, local outlets, podcasts, the right review sites. A niche blog with five thousand of the right subscribers can move more than one trophy hit, because the machine is counting agreement, not audience.

No single page is the needle. The consistency across your site, your records, and other people’s coverage is. If you are a solo founder or an artist with no press at all, the smallest version of this still works: name and one-line description identical in three places, an FAQ on your About page, and one run of the audit. Start there.

What should you stop paying for?

Everything on this list is a shortcut that tries to skip the work, and the machine does not bite.

Stop chasing the old scoreboard. The Domain Authority number first. Across 21,767 domains, every classic authority score (Moz DA, Ahrefs DR, Domain Power) correlated near zero or slightly negative with AI visibility, between −0.08 and −0.21. Note what I am saying and what I am not: stop chasing the number, not stop earning coverage. The work that raises DA is the same third-party vouching that AEO rewards. Wrong scoreboard, right gym.

Then llms.txt, the proposed file you are supposed to drop on your site to instruct the models. SE Ranking tested nearly 300,000 domains and found no correlation with AI citations; their prediction model actually improved once they removed the variable. Google’s own people still call it “purely speculative.” Do not confuse it with robots.txt, which is real and matters enormously.

And FAQ schema, the hidden code version. Google fully retired FAQ rich results on May 7, 2026, and Ahrefs ran a controlled study of 1,885 pages that added schema against roughly 4,000 controls and found no detectable lift in AI citations. Build the visible FAQ for humans. Skip the hidden markup for robots.

Stop buying press before you are an entity. There is a clean natural experiment on this. A UK company planted 11 fabricated experts across 600+ real articles in the British press, Press Gazette exposed it, and Authoritas then tested how nine AI models handled them. On the 55 topic-based questions, the kind a real customer types, the fake experts surfaced zero times. Worth stating the limit honestly: when researchers asked about the fabricated names directly, models validated them about 29.5% of the time. So the machines were not perfect at detecting fakes. They were just completely uninterested in recommending them to anyone who had not already heard the name. That is hundreds of real placements multiplied by an entity value of zero.

Press amplifies an entity that exists. It cannot create one.

Stop renting black boxes. Agencies and tools that cannot show you their validation data. Not because tools are evil and not because every agency is useless, but because the field is two years old and mostly cannot measure itself yet. Ask any vendor for the data behind their accuracy claim before you pay. Most cannot produce it. If you are going to spend money on a human, the highest-leverage version is usually one niche PR consultant with a decade of real relationships in your exact world, rather than a shop running the same playbook across twelve clients.

The meta-stop underneath all three: stop looking for a trick at all. Every item here is somebody selling a way around being good.

What does this actually come down to?

There is a thing in the old SEO world where you ask a serious technical practitioner, a Gianluca Fiorelli type, what to do first, and half the time the answer is “compress your images.” The least glamorous instruction imaginable, and it works. AEO runs on the same energy. Fix your entity. Get your name consistent. Show up in the dull databases nobody brags about. The unsexy work is the work, and nobody claps when you fix your Crunchbase.

Which is also why I find this more encouraging than the thing it replaced. You cannot talk the machine into trusting you, and you cannot fake corroboration across every surface at once. What you can do is give the world something true and let other people repeat it. When we wanted to be trusted, we ran a study on how often promo codes actually fail at checkout and published the result even though it was unflattering to a coupon company, and it got picked up across something like 150 media placements in over a hundred local TV markets. We never said we were credible. A hundred other sources did.

There is less room for smoke and mirrors now than there was five years ago. The name of the game is closer to actually being good than it has been in my whole career, which is either wonderful or terrifying depending on the week.

So: run the audit before you change anything, because you cannot fix what you have not seen. Then pick the one gate you are failing and work on that alone until it moves.

It will feel slower than you want. Entity work is measured in months, because you are waiting for several independent sources to agree about you, and there is no version of this where a good week shows up as a spike. What you get in exchange is the only kind of visibility that survives the next model update.

The audit is the companion to this piece: The 30-Minute AI Visibility Audit is the free method, written out step by step. If you want the argument for why any of this is worth your Monday, Good Content Strategist vs. Bad Content Strategist covers the posture underneath it.