Bill Hunt tested 190 AI answers, 74 percent lacked real evidence

Bill Hunt tested 38 buyer prompts across five AI tools for one company. He got 190 answers in total. The company was mentioned in 98.4 percent of them. It was the primary pick or shortlisted in 93.2 percent. Hunt then checked what the answers said behind the mention. 73.7 percent had weak or missing evidence for the key claim. The piece is in Search Engine Journal and is dated 11 November 2026.

What Hunt measured

The prompts are what buyers would actually ask, not brand searches. They cover workflow fit, integrations, and rules the buyer must follow. They also cover how the product works with other tools, and what the company reports back to the customer. Hunt ran the same 38 prompts on five AI tools. He names ChatGPT, Gemini, Claude, Perplexity, and Google.

That gave him 190 single answers for one company. Hunt did not run the same prompt many times. One answer per prompt per tool.

The two numbers that matter

The first set of numbers is the visibility: 98.4 percent mention rate, 93.2 percent on a shortlist or named as the top pick. That is the number most GEO tools report. Hunt says it looks like a good scoreboard.

The second set is where the story is. When a buyer asked a second, sharper question, the answers fell down. 73.7 percent had weak or missing evidence for the main claim. 15.8 percent had critical gaps on how the product fits the buyer's work.

By Hunt's account, the company often had the feature and the proof. The information was broken into pieces across the site, old blog posts, sales decks, and a partner's page. An AI answer could not ground the claim from any single page.

The fix Hunt recommends

Hunt argues the people who need to add evidence are product, engineering, sales, compliance, and the company's partners. Each one owns a proof the model needs. Product owns a feature that works as described. Engineering owns a workflow that matches the buyer's. Compliance owns a rule the product passes. Sales owns a customer who is willing to be named. Marketing alone cannot write any of these.

Hunt recommends building a scorecard that matches real buyer criteria, and scoring the company's content against it. Each claim is rated on whether a third party backs it. A brand saying something about itself does not count. Measure buyer progression through a chain of prompts, with more than one answer.

Where the method stops

The study is one company, one researcher. Hunt does not name the company. He does not say what the product is. A reader cannot check his scoring against the live AI answers the way a reader could against a vendor benchmark. The method is useful as a lens for a team to run against its own site. It is not a general number for the industry.

The percentages above describe the single company in the study. Read them that way. Hunt's larger argument is that visibility and evidence are two separate scores. A team watching only the first one is measuring the easy half.

Source

Bill Hunt, Search Engine Journal.