How AI Agents Choose Businesses.

The decision criteria AI systems use when recommending a firm are specific, observable, and mostly orthogonal to what the traditional marketing stack optimizes for. Here is what they actually weight — and why "ranking well on Google" does not translate.

Ask ChatGPT "who's the best trusts and estates attorney on the Upper East Side?" Ask Perplexity "best fee-only RIA in Westchester for concentrated stock positions?" Ask Claude "top cyber liability broker for fintech companies?" The systems return answers. Specific ones. Named firms. Paraphrased positioning. Occasional citations.

Those answers are not generated by running a Google search. They are generated by assembling signals the model has learned, from its training data and its live retrieval, to weight heavily in category decisions. Understanding which signals are being weighted is most of the battle.

Signal one: third-party citation density

The most predictive variable for appearing in an AI answer is the density and quality of third-party sources that name your firm in your category. Not the quality of your own website. Not the depth of your own content. The number and authority of other people's pages that describe you as a firm in your category.

This is not a small distinction. A law firm with a beautiful website and no meaningful third-party presence will be invisible to AI for the exact queries it most wants to win. A law firm with an ugly website and robust citations in the ABA Journal, the Law360 archives, state bar publications, and reputable local journalism will be named readily.

The implication: when you are investing in GEO, a large share of the work lives off your own property.

Signal two: credential verifiability

AI systems are conservative about recommending firms in regulated industries unless they can match a claim on the firm's site to an authoritative public record. A "board-certified surgeon" claim on the firm's site means very little. A Physician schema block, paired with an NPI record, paired with a hospital affiliation page, paired with a specialty society member page, means a great deal.

For law, the equivalent chain is: bio claim → state bar admission record → law-school faculty / court record presence → authored content attributable to the named lawyer.

For wealth: bio claim → FINRA/IAPD record → CFP Board or CFA Institute registry → authored content.

The pattern in every vertical is the same: the AI is not taking your word for it, and structuring your site so the match is cheap to make is a first-order act of visibility engineering.

Signal three: extractable content patterns

AI systems don't cite paragraphs they cannot cleanly extract. Long, narrative, persuasive marketing copy — the kind most professional services sites are full of — is not extractable. It reads well to humans and extracts badly.

The patterns that extract cleanly are consistent across platforms: direct answer in the first sentence, 140–170 word paragraphs built around a single claim, verifiable facts every 150–200 words, heading hierarchy that a parser can use, and schema that identifies what each section is about.

The firms that read as "thoughtful content marketing" to a human and "extractable paragraphs" to a parser are the ones that get paraphrased into the answer. The firms that read well but extract poorly get passed over.

Signal four: recency and update cadence

Retrieval-heavy AI systems — Perplexity, Google AIO, and increasingly ChatGPT in its web modes — weight recency more aggressively than most SEO professionals expect. A firm that last updated its service pages in 2022 is competing against firms that updated last month. In AI terms, the 2022 page is closer to invisible than it looks.

Update cadence is not about pretending to publish. It is about ensuring that when an AI retrieves content in your category, your firm's pages are among the ones that look alive.

Signal five: co-occurrence with category peers

AI systems use co-occurrence patterns to learn category membership. If your firm is named alongside the top five firms in your category on independent pages — a trade publication's list, a ranking article, a conference speaker lineup — the model learns that you are a peer of those firms.

If you are never named alongside them, you are, in the model's representation, not in the category.

This is one of the least appreciated dynamics in GEO. A single well-placed mention as a peer of the category leaders often moves citation share more than ten pages of on-site content.

What this means for investment priorities

Most professional services firms invest in the wrong layer for AI visibility. They pour resources into on-site content quality — useful but insufficient — while under-investing in the third-party, credential, and co-occurrence signals that actually shift AI recommendations.

On-site content work still matters — it is part of every engagement we run. But it is not the lever most firms think it is. The lever is usually somewhere else in the stack, and the point of measurement-first GEO is to find out where.