What do AI engines look for when they recommend a company?

Last updated October 14, 2026. Written by the Third Coast AI team.

When a buyer asks an AI engine who to hire, the engine treats the question as a small request for proposals: it decides what the criteria are for that category, runs a set of hidden searches to gather candidates and evidence, scores what it can read against those criteria, and writes a recommendation, which means the companies that get named are the ones whose own pages and third-party record supply proof of the specific decision factors the engine is using, in words a machine can quote. The factors are different for every category and they are discoverable, because the engines will tell you what they are if you ask them the right way.

This page explains the mini-RFP model, shows how to extract the decision factors for your category from the engines themselves, and describes how to audit your site for proof of each one. It is the step in our method that most companies skip and the one that produces the most change per hour of work.

Why is a buyer's question a mini-RFP?

Because the engine does not search for the words the buyer typed. Google's AI features page says its AI systems may issue multiple related searches across subtopics and data sources to develop a response, and our own captures routinely show 10 to 20 hidden sub-queries behind one buyer question. Ask who should build a pool for a backyard in Ottawa County and the engine goes looking for licensing, years in business, gunite versus fiberglass, warranty terms, financing, recent reviews, permit handling and nearby alternatives, then assembles an answer from whichever pages answered those. Each sub-query is, in effect, a line item on an RFP, and a company whose site answers four of twelve line items scores four of twelve no matter how well written the home page is.

The scoring is not a formula anyone outside the labs can see, and it is not the same in every engine; only about 11% of cited domains overlap between ChatGPT and Perplexity. But the shape is consistent. Zhang, He and Yao's 2026 study of 21,143 citations found the pages with the most influence on the written answer were the ones rich in extractable evidence such as definitions, facts, comparisons and procedural steps, which is a description of a well-answered RFP. And Cyrus Shepard's May 2026 synthesis of 54 studies scores crawler accessibility at 9.5 out of 10 and search rank at 9.4, which are the two things that decide whether your proposal is even opened.

How do you find the factors for your category?

1. Write the prompts the way buyers ask. Not keywords with the word best in front, but the sentences a real buyer types, including the awkward ones: whether to pay more for a local firm, what happens if the project runs late, whether the company handles a specific edge case.
2. Ask the engines for their criteria directly. Prompts of the form what should I look for in a [category], how do I compare [category] companies, and what questions should I ask a [category] before hiring one, run in each engine with retrieval on and more than once.
3. Capture the hidden searches. Where an engine shows the sub-queries it ran, record them; where it does not, infer them from the cited pages. The sub-queries are the factors in search form.
4. Count and rank. List every criterion the engines returned, in their words, and count how often each appears across engines and runs. The top ten are your RFP.
5. Read the cited pages. Open every page the engines cited for the factor prompts and note which factors each one states and proves. That is the standard you are being scored against.

The factor list is usually a surprise in one of two directions. Owners expect price to lead and it rarely does; it sits mid-list behind proof of competence and specificity to the buyer's situation. And owners expect their differentiators to appear and they often do not, because a differentiator the engine never asks about is not a factor, however true it is. Peec's March 2026 analysis of 30 million sources found ChatGPT adds the word reviews to its own hidden searches even when the user did not ask, which is the clearest example we know of a factor nobody listed on their site because nobody knew it was being scored.

Which proof does the engine accept?

Proof the engine can quote. For each factor, the question is whether there is a sentence on your site that states it plainly and backs it with a number, a named source or a verifiable claim, and whether anyone else on the web says the same thing. Licensing is a license number and the issuing body. Years in business is a founding year that matches Wikidata and the Google profile. Response time is a stated commitment with a mechanism behind it. Certifications are named with the certifying organization. Reviews with recency are a review profile you keep current, on the platform the engine cites for your category. Ahrefs' 75,000-brand study found third-party mentions correlated with AI citations at 0.664 against 0.218 for backlinks, and the practical consequence for factors is that a claim only you make is a claim the engine discounts, while a claim your site makes and a trade publication repeats is one it relays.

Can you audit your own site for this?

Yes, with a spreadsheet. Put the factors down the side and your pages across the top, and in each cell write stated, proven or missing. Most companies find they state about a third of the factors and prove fewer, and the cells marked missing are the brief for the next month's pages, in the shape our structure guide describes. The second pass is off-site: for each proven factor, is there a third-party page that repeats it, and if not, which of the sites the engines cite for your category could. That list goes to whoever handles outreach, and our mentions page explains why it outranks link building.

Do the factors change?

Slowly, and in two ways. The engines update their source mix and their models, so a factor that leaned on one review platform can shift when that platform's share of citations moves, which is why we re-run the factor prompts quarterly. And your competitors publish, so a factor that nobody proved last quarter can become one that two companies prove this quarter, which raises the bar for the third. The audit delivers the factor list, the proof audit and the cited pages for one category in three weeks, and the proof audit is the part most clients keep using after the engagement ends.

Frequently asked questions