I'm a Computer Scientist. Here's How AI Search Actually Decides Which Company It Recommends

Most articles about ranking in AI search are written by marketers guessing at a black box. I am not guessing. I spent fifteen years building software, including AI and machine-learning work that goes back to a Bayesian inference project I published on in college, and engineering on federal systems before I ever ran a marketing team. So when a buyer types "best company for X" into ChatGPT and your business does not appear, I can tell you what actually happened, not what a content calendar says happened.

Here is the part nobody wants to admit. The model did not decide you were worse than the company it named. In most cases it never saw you at all.

The answer is assembled, not recalled

When someone asks an AI tool to recommend a company, the system is not reaching into a tidy memory of every business in your category. Large language models do not store a clean directory. They store statistical patterns about language. Ask one for a specific company by name and you will often get a confident, wrong answer, because the model is predicting plausible text, not reading a database.

So the good tools cheat. ChatGPT with browsing, Perplexity, and Google's AI answers do not respond from memory. They run a real-time retrieval step first. They issue searches, pull back a handful of sources, and then write an answer grounded in those sources. The language model is the writer. The retrieval layer is the bouncer at the door, and it decides who gets into the room before the model writes a single word.

That distinction is everything. There is what a model can dredge up from training, which is recall, and there is what it goes and fetches in the moment, which is retrieval. Recall is a party trick. Retrieval is the one you can influence, and if you are not in the retrieved set, no amount of "quality" gets you into the answer. You lost before the model started typing.

Four levers actually move you into the retrieved set

Once you see it as a retrieval problem, the levers stop being mysterious. Here is where I focus, in order of impact.

1. Machine-readable identity. The retrieval layer needs to parse what you do, who you serve, and where you operate, in a structured form it does not have to interpret. That means schema markup. Organization, Product, Service, and FAQ schema on your site, with a consistent name, location, and category, is the difference between a crawler understanding "B2B logistics software for mid-market manufacturers" and seeing an unlabeled blob of marketing copy. Most sites I audit have none of this, or have it broken. The competitors who show up usually have it clean.

2. Corroboration across independent sources. Retrieval systems weight agreement. When your name, your category, and what you sell say the same thing on your site, your LinkedIn page, your Crunchbase or industry directory listing, a review platform, and a press mention, the system gains confidence that you are a real, specific entity worth surfacing. When those sources disagree, an old tagline here, a former company name there, a wrong location somewhere else, the system discounts you to avoid recommending something wrong. Consistency is not housekeeping. It is a ranking signal.

3. Content that answers the actual question. The queries that matter are not "logistics software." They are "how do I cut freight costs without switching carriers" and "what is the fastest way to onboard a new distribution site." Those are the questions buyers ask the AI, and the AI pulls its answer from pages that address them directly. A site that is all product pages and no substance feeds nothing into that pipeline. A site with clear, specific, genuinely useful answers becomes a source the model quotes. You want to be the page it cites, because the citation is the new front door.

4. Freshness and crawlability. Retrieval favors pages it can reach and pages that look maintained. If your site blocks crawlers, buries its content behind scripts the bots do not execute, or has not published anything since 2022, you are a stale source. Stale sources get passed over for active ones. A machine cannot cite what it cannot load, and it would rather not cite what looks abandoned.

What does not move the needle as much as you think

Your star rating matters to humans. It matters much less to the retrieval step, because the model is not running a popularity contest. I have watched businesses with hundreds of glowing reviews stay invisible while a smaller competitor with clean schema and a real content library got named in the answer. Reviews help you close the buyer once they find you. They do very little to get you found in the first place.

Ad spend does not help here either. The AI answer is not an auction. You cannot buy your way into the cited sources. You earn your place by being the most parseable, corroborated, relevant source the retrieval layer can find. That is a different game than the one most marketing budgets were built to play.

Why this is moving fast

The share of buyers who open an AI tool instead of a search engine is climbing every quarter, and it climbs fastest among exactly the people you want, the ones who research carefully before they ever fill out a form or take a call. The larger, better-funded players in your category already know this. They have the engineering budgets to fix their structured data and publish at volume. The operator who treats AI search as next year's problem is handing those competitors a head start that compounds.

The good news for a lean company is that this is a technical problem, and technical problems have concrete fixes. You do not need a bigger ad budget. You need your site to be legible to a machine, your identity to agree with itself across the web, and a library of pages that answer what buyers actually ask. That is buildable, and it is buildable faster than most people expect.

How to check where you stand

Open ChatGPT or Perplexity and ask it, the way a buyer would, for the best company in your category and market. Then ask the follow-up questions a real buyer asks. See whether you appear, see who does, and look hard at what those companies have that you do not. Nine times out of ten the ones who show up have cleaner structured data and more on-topic content, not a better product.

That five-minute test is the front end of what I call the Machine Score, the audit I run to grade how well AI search can find, parse, and trust a business, and to hand back exactly what is missing and what it would take to fix. You can start it yourself today with nothing but a chat window and an honest look at the results. If the machine cannot find you, that is where the work begins.

William Hunt
Hunt and Machine | ryan@huntandmachine.com