Engine deep-dive

How ChatGPT decides what to recommend

OpenAI has published nothing about how ChatGPT selects which companies to recommend. Everything in this article draws on published industry analyses of AI search citations and how retrieval works. Updated June 2026.

Two ChatGPTs, two recommendation systems

This is the most important thing to understand. ChatGPT operates in two distinct modes, and each produces different recommendations for the same query:

Most users now have browsing enabled by default. This means your content strategy can directly influence ChatGPT recommendations, often within days of publishing. But the base model still serves a significant portion of queries, especially in the API (which many tools and agents use).

The 5 signals that matter

Drawing on published industry analyses of AI search citations across B2B service categories, five signals consistently correlate with being recommended:

What does NOT matter

• Google search ranking. A company can rank number 1 on Google for its primary keywords and still be absent from ChatGPT responses for the same queries. The systems are independent.
• Paid advertising. There is no paid placement in ChatGPT. Ad spend does not buy AI recommendations.
• Domain authority. Traditional SEO metrics like DA/DR do not predict ChatGPT citations. A small company with a specific, well-structured page can outrank a Fortune 500 with a generic one.
• Social media followers. Follower counts on Twitter, Instagram, or Facebook do not appear to drive ChatGPT recommendations.

The white space opportunity

For many B2B service queries, ChatGPT gives generic, non-committal answers. It frequently says "there are several good options" without naming a specific company. That gap is the opportunity.

Ask "best MSP for healthcare compliance" and ChatGPT often lists well-known general IT companies but no MSP that actually specializes in healthcare. The specific query has no specific answer. The first MSP to publish a comprehensive, schema-marked page about healthcare IT compliance is well positioned to own that recommendation.

The lead time is real but short. Search-grounded retrieval re-reads the live web, so structured-data changes can show up in cited answers within weeks, not quarters.

What you can do today

1. Add FAQ schema to your top service pages. 5 questions matching the queries your buyers ask ChatGPT. Under 80 words per answer. This is the single highest-ROI action.
2. Create one vertical-specific page. "[Your Service] for [Specific Industry]." 2,000+ words. Include FAQ schema. Target a query where ChatGPT currently gives a generic answer.
3. Get on Clutch. Create or update your Clutch profile. Ask 5 clients for reviews. Clutch is a frequently cited B2B review source in ChatGPT responses for consulting and services queries.
4. Check your robots.txt. Make sure GPTBot and ChatGPT-User are not blocked. If they are, ChatGPT cannot crawl your site in browsing mode.

What our own scans measured

We can put a number on how often ChatGPT names anybody at all. Across 46 completed scans on our own platform it named the brand in 1.7% of the buying questions we put to it, and every firm in that set had run a scan because someone suspected it was missing, so this is a rate for firms with a suspicion and not a random sample of B2B websites. That measurement used the first of the two modes described above, the one that answers from training weights without searching, because that is what the API does unless you force it otherwise. We began forcing web search on 2026-08-18. So read that number as what ChatGPT remembered, not as what ChatGPT can find, and expect it to move once we have enough scans on the other side of that change to state a new one honestly.

The strongest engine on the same reports was Perplexity at 7.4%, with the rest of the field between 0.4% and 4.0%, and the same caveat applies because these are the same firms that suspected a problem before they scanned. Absence is the ordinary condition. A case study claiming your category is already saturated is describing a handful of questions, not the general state of things.

A second measurement matters more here than the headline rate. In July we asked four engines to recommend managed IT providers across 15 North American regions and resolved every name they produced to a live website: 8 of the 114 names ChatGPT gave us could be corroborated. Perplexity, which searches before it answers, corroborated 101 of the 257 names it produced in that same managed IT fieldwork. A model drawing on memory will always produce a name, and a name is not always a company.

Read the MSP AI Visibility Report 2026, which shows the method and the names we could not verify.

How we count, because the denominator does the work

These rates cover buying questions only, the ones a buyer types when they are choosing a vendor and do not yet know who you are. We leave out questions that contain the company's own name, deliberately. Across the same 46 reports every engine answers a name-in-the-question probe with that company 82% to 100% of the time, so folding those in lifts the headline to about 10% for the field and 19% for the leader. We published that flattering version on this page until 2026-08-12 and have corrected it. The buyer who already knows your name is not the buyer you are missing, and a number that counts them is measuring your own marketing back at you.

What you get, and what it costs

ChatGPT is one of the 6 engines every plan queries, including the free one. Engines are not plan-gated here. What a plan changes is how many questions you get per scan, how often the scan runs, and whether we draft the pages that close the gaps it finds.

Rule is portfolio level and sales-assisted. Full pricing · How the scan works