How AI assistants choose which B2B vendors to recommend

A language model answering without search relies on patterns learned during training. It can name vendors that were widely discussed in its training data, but it cannot know about your launch last quarter, and it may describe you as you were years ago. That memory changes only when models are retrained.

Most buyer-facing assistants now ground commercial answers in search. Google's AI Overviews and AI Mode build answers from Google Search results. Microsoft 365 Copilot fetches information from the Bing search service when web results help. OpenAI says ChatGPT search surfaces public websites that do not block OAI-SearchBot. Perplexity searches its own index for every answer. Retrieval is where marketing work shows up within weeks rather than model generations.

The two sources interact. A vendor the model already associates with a category is easier to name, and retrieved pages can confirm, update or contradict that memory. That is why fixing your own pages is necessary but not always sufficient: if most of the web still describes you the old way, the answer may too.

The details differ by engine, but grounded answers follow the same broad sequence.

1. The buyer asks a long question, often with constraints: industry, company size, existing systems, region, budget.
2. The assistant rewrites it into one or more search queries. Google describes AI Mode breaking a question into subtopics and issuing many queries at once; Microsoft says Copilot generates a query of a few words that differs from the prompt.
3. The search system returns candidate pages, and some engines rank passages within them. Perplexity says it retrieves and ranks self-contained spans, not only whole pages.
4. The model reads the top results and writes an answer, naming the vendors those sources support for the question.
5. The interface shows citations to some of the sources used, which is where you can see why a vendor was named.

No engine publishes a ranking formula for vendor recommendations. From the official descriptions and the better studies, five properties keep coming up.

Notice what is missing: ad spend, schema volume and llms.txt files. Google states it uses neither special schema nor llms.txt for its generative AI features.

Rand Fishkin of SparkToro and Patrick O'Donnell of Gumshoe.ai had 600 volunteers run 12 identical prompts in ChatGPT, Claude and Google's AI, nearly 3,000 runs in total. The odds of getting the same list twice were under 1 in 100, and the odds of the same list in the same order were closer to 1 in 1,000. What was stable was presence: some brands appeared in 60% to 90% of responses.

That changes how to report. A screenshot showing you third, or missing, says almost nothing. The useful metric is visibility rate: across many runs of the same buyer intent, how often are you named, and how often are you named correctly. Our guide to measuring share of answer sets out the method.

Buyers increasingly start here. Gartner reported in March 2026 that 45% of B2B buyers used AI during a recent purchase, and that 67% prefer a rep-free experience. G2's 2026 survey, which is vendor research, found 69% of software buyers chose a different vendor than planned because of a chatbot recommendation and that one in three bought from a vendor they had never heard of before.

For a B2B company, that means the AI shortlist works like a first-round filter you never see. Our article on how B2B buyers research vendors in ChatGPT covers what happens after the answer.

Each property above maps to work a marketing team can own.

• Retrievable: audit crawler access for Googlebot, Bingbot, OAI-SearchBot and PerplexityBot, and fix rendering and indexing gaps.
• Associated: use the same category language on your site, profiles and partner pages that buyers use in prompts.
• Confirmed: earn mentions on the domains AI answers already cite for your category.
• Specific: publish pages for the constraints you win on (industries, integrations, company sizes, regions) with the answer in the first lines.
• Current: review the pages that carry your visibility every quarter and update facts when they change.

For a week-by-week version of this work, see our 30-day plan to get recommended by ChatGPT.

Build the prompt set from real buyer language, run it enough times to see frequencies, and record the cited sources. That baseline tells you which of the five properties is your bottleneck.

Our generative engine optimization service runs that baseline and the fixes across Google, Copilot, ChatGPT and Perplexity.

If you want to know why assistants recommend your competitors, talk to our team.