Where AI Gets Its Answers About Vendors

Buyers now ask AI assistants which vendors belong on the list. The answer gets assembled from pages the vendor did not write, and few vendors have checked which pages those are.

Shortlister’s marketplace lists more than 8,000 benefits vendors across 290-plus categories, and nearly all of them maintain their own websites.

Ask an AI assistant about one of those categories, though, and the answer pulls little from those sites. Vendor discovery today, starts with someone else’s description of the company. Which pages are those, and why do they win the citation?

The Question a Model Has to Answer, and Why a Homepage Can’t

Renewal season brings many questions like this one: Which EAP vendors suit a 1,500-person manufacturer with a self-funded plan?

Answering it well means naming several EAP providers and comparing them on features the buyer cares about, such as pricing and the populations they serve. No single vendor site can supply that, however strong its copy is, because it says nothing about the competition.

So, the AI tool must search elsewhere.

As Shortlister’s earlier piece described, most assistants split one prompt into several narrower searches, an approach Google calls “query fan-out.”

An AirOps study of 548,534 pages retrieved across 15,000 prompts found that 32.9% of cited pages surfaced only through those follow-up searches. Yet 85% of everything retrieved never made it into an answer.

A vendor’s homepage can end up in that retrieved pile and still miss the citation to a page that answers the whole comparison at once.

In short, being retrieved isn’t the same as being cited.

The Four Kinds of Sources That Get Retrieved for Category Questions

The sources that tend to appear fall into four broad types: comparison pages, trade press, community forums, and the vendor’s own site.

1) Category and Comparison Pages

Directories and category lists answer a comparison question in one place, which is why AI models reach them early.

Wix Studio’s AI Search Lab analyzed more than 1 million citations and found that listicles took 40.9% of citations for commercial-intent queries.

In professional services, third-party lists accounted for 80.9% of those listicle citations, against 19.1% for lists vendors published about themselves.

In other words, a vendor’s own “best of” page rarely wins the citation when an independent source covers the same topic.

Where this format falls flat is depth and currency.

Plan-year changes, new carrier connections, and updated certifications may not reach a directory at all, or may only be reflected late. Until they do, the model can only retrieve the old version.

2) Analyst Notes and Trade Press

Editorial sources get picked up when a question depends on recent events.

Semrush’s three-month tracking showed Forbes roughly doubled the number of ChatGPT responses citing it after a Google search change in September 2025.

However, benefits vendors rarely gain from that.

A regional voluntary benefits provider is far less likely to receive the same press attention as a national carrier, leaving smaller vendors with much less third-party material for AI systems to retrieve.

3) Practitioner Communities and Forums

Forums such as Reddit and Quora offer something unique: firsthand accounts of implementation and service.

SE Ranking’s study of ChatGPT citation factors found that domains with millions of brand mentions on Reddit and Quora were cited about four times as often as domains with little activity there.

However, Reddit’s and Quora’s visibility in AI answers can change rapidly, depending on how each AI platform adjusts its retrieval.

According to Semrush, ChatGPT cited Reddit in close to 60% of responses in early August 2025 and only around 10% by mid-September, and Quora lost ground in Google’s AI Mode over the same stretch.

For B2B evaluation questions, however, these communities tend to be less active and offer fewer useful discussions for buyers.

Ten Speed tracked evaluation-stage B2B prompts and found community sources only made up around 4% of citations.

Forums can, to a degree, shape how buyers talk about a vendor, but they rarely decide whether an AI assistant recommends one.

4) The Vendor’s Own Site and the Questions Where It Does Win

The go-to retrieval source changes once a buyer moves from discovery to evaluation.

Once a buyer names specific vendors and requests an evaluation, the AI model needs facts only the vendors can state with authority.

Across 170 evaluation-stage B2B prompts, Ten Speed found product pages drew 24.1% of citations, the largest single format, and brand-controlled content made up 88% of what was cited.

Homepages took about 8%, which the study interpreted as models checking who a company is and what it sells before answering a narrower question.

Category questions, then, go to pages that cover many vendors, while specific questions about a named company go to that company’s site.

What’s crucial is that a vendor must be described accurately in both places.

What Makes a Category Source Machine-Readable

What makes one source easier for a model to compare vendors across? Mostly its structure.

A vendor website can describe its own offering in whatever language and format it chooses. A category directory can organize many vendors around the same questions, fields, and terminology, giving a model almost a complete dataset it can compare rather than simply retrieve.

Four properties make that difference.

A Consistent Taxonomy

Vendors are classified using the same category names and product definitions. If one company calls its offering “care navigation” and another calls a similar service “health advocacy,” a model has to work out whether they belong in the same comparison. A shared taxonomy removes much of that ambiguity.

Missing Capabilities Recorded

Vendor sites advertise what they offer and almost never explicitly list what they do not. A category source records “not offered” in the field, so a model treats a gap as a fact rather than guessing whether the vendor skipped it.

Fit Criteria Recorded Consistently

A vendor’s website may mention things like minimum group size or availability somewhere in its copy, but that information is not always easy to find or compare.

A structured directory can record those details in the same fields for every vendor. That helps a model quickly see which vendors fit the employer’s requirements and which do not.

Broad Category Coverage

A source that describes a substantial share of the vendors in a category gives the model a comparison set in one place. Search specialists call that breadth topical authority, and it lets a model see which capabilities are common and where vendors differ instead of piecing together isolated claims from company sites.

Why This Changes What a Vendor Profile Is For

For most of their history, directory listings were built to be found and read by people. Yellow Pages advertisers named their companies “AAA Plumbing” to rank first in their section, and early directory-listing SEO carried the habit online with keyword-stuffed profiles.

In a structured marketplace, though, a vendor profile now has a second reader – AI models. Those models can use profile information to answer a buyer’s question without requiring the buyer to open the page.

That changes the role of a vendor’s profile because buyers increasingly get answers without visiting every source behind them.

In fact, Pew’s browsing data found that people who saw a Google AI summary clicked a traditional search result on 8% of visits, compared with 15% when no summary appeared. Only 1% clicked a link within the summary itself.

In other words, the description surfaced during research may influence the buyer before the vendor ever gets a site visit.

That early impression matters increasingly more in B2B buying.

Bain and LinkedIn found that around 90% of buyers ultimately purchase from their day-one vendor list. Shortlister’s own analysis in The Shrinking Shortlist shows that invitation lists have narrowed sharply since 2022, leaving less room for vendors to enter the process later.

If structured profile data can shape how a vendor is classified, compared, and surfaced during early research, maintaining it becomes an operating task with a named owner and a review schedule.

How to Check Where a Model Got Its Answer

The check takes about 20 minutes and includes a spreadsheet. It records which sources an assistant relied on, something a count of vendor mentions doesn’t reveal.

Write four prompts in a buyer’s voice: two broad category questions (“Which telehealth vendors should a 2,000-employee retailer consider?”) and two with a specific requirement (“Which fertility benefits vendors include both IVF support and adoption assistance?”).
Run each prompt in ChatGPT, Gemini, Perplexity, and Claude with web search turned on.
Follow each answer with “List every source used for that answer, with URLs.”
Log the date, assistant, prompt, vendors named, cited URLs, and whether the company description is accurate.
Open each cited URL and verify that it exists and states what the answer claims.
Repeat on the first business day of each month, using the same wording.

Running all four assistants matters because their sourcing differs. Studies have found that Claude and ChatGPT relied most heavily on earned sources, while Perplexity and Gemini drew more on brand and social content.

Written by Ivana Radevska

Senior Content Writer at Shortlister

Sources

EAP Providers

Browse our curated list of vendors to find the best solution for your needs.

Stay Informed

Subscribe to our newsletter for the latest trends, expert tips, and workplace insights!