The B2B shortlist is built inside the AI chatbot now: how buyers choose software before they see your site

Half of B2B software buyers now build their shortlist inside a chatbot before they open a vendor's website. A March 2026 survey of 1,076 buyers found 51% now start research in an AI chatbot more often than Google G2 · 2026B. That share was 29% a year earlier, and 69% chose a different vendor than they had planned on the chatbot's guidance G2 · 2026B. The shortlist is short, it favours brands the web already describes, and a review-site citation is the proof buyers look for before they trust it. What earns a slot is retrievable, consistent facts about the product, corroborated off-site. The sections that follow separate the behaviour shift from the advice being sold on the back of it.

The shortlist forms before a buyer reaches your site

The first impression of a vendor is now formed by what a chatbot says, not by the vendor's homepage. In G2's survey, 51% of buyers start research in an AI chatbot more often than Google G2 · 2026B. 71% use one at some point in the process G2 · 2026B. The move from search to synthesis changes who gets considered. Buyers ask the assistant to return the best options rather than a list of links, so the assistant curates the shortlist.

That curation decides deals. 69% of buyers chose a different vendor than first planned, on a chatbot's recommendation G2 · 2026B. A third bought from a vendor they had never previously heard of G2 · 2026B. This is a single-vendor survey from a company that sells answer-engine products, so treat the exact figures as directional. The wider zero-click shift corroborates the direction: B2B click-through rates fell as much as 30% after Google introduced AI summaries Bain & Company · 2025A. Fewer buyers click through to judge for themselves.

The shortlist is short, and it favours brands the web already describes

An AI shortlist names a handful of brands, not a page of them. Across categories, AI assistants recommend an average of 6 to 11 brands per prompt, and established leaders dominate the answers in many sectors Similarweb · 2026A. Being left off a list that short is expensive, because the buyer often does not see who was excluded.

What predicts inclusion is how much the rest of the web talks about a brand. Across 75,000 brands, branded web mentions correlated with AI Overview visibility at 0.664, against 0.218 for branded backlinks, roughly three times the association Linehan · Ahrefs · 2025A. The signal that moves the shortlist is being described elsewhere, in the sources the model reads when it assembles an answer. A brand that no third party describes has little for the model to weigh.

Recognition is not recommendation

A brand can be known to the model and still never make the shortlist. AI assistants recognise a named product almost universally, but recommend it in open discovery queries far less often Sharma · arXiv · 2026A. The gap between the two is the right test of AI visibility Sharma · arXiv · 2026A. Typing the company name into a chatbot and getting a clean description measures recognition. It does not measure whether the model would name the brand unprompted.

Referring domains, not on-page optimisation scores, predicted whether a product surfaced in those discovery queries Sharma · arXiv · 2026A. The buyer behaviour and the retrieval mechanics point the same way. The buyer asks an open question, and the assistant answers from the brands the web describes and links. A brand recognised only when the buyer names it is not in that set.

A review-site citation is the proof buyers want, and it corroborates a story the model already holds

Buyers accept an AI recommendation, then look for evidence behind it. The trust signal that most increases buyer confidence in a chatbot's answer is a citation from a software review site G2 · 2026B. The reading that review volume alone wins the shortlist goes too far. A citation confirms a recommendation the model has already made from the facts and mentions it could retrieve.

Third-party presence does move inclusion, which is why the review layer matters. Across 5.7 million brand-mention observations, appearing in repeatedly cited third-party listicles raised a brand's odds of being mentioned by 14 to 32 percentage points Ehrlinspiel et al. · Peec AI · 2026B. The study separates being selected from being made prominent, and both are earned off the brand's own site. Reviews and third-party coverage are the corroboration layer. They sit on top of the facts the model has already gathered, rather than replacing them.

What earns the slot is retrievable, consistent facts, not markup

The shortlist is built from facts the model can read. The lever is the accuracy and consistency of those facts, not the code around them. Adding schema to 1,885 pages moved AI citations +2.4% on Google AI Mode and +2.2% on ChatGPT, both indistinguishable from zero Ahrefs · 2026B. Markup is hygiene. It does not put a brand on the shortlist.

For commercial questions, the answer is built mostly from the brand's own pages. Across 127,198 citations on commercial queries, 90.6% pointed to vendor, product, or documentation pages rather than to forums or encyclopaedias Khallad · SurfacedBy · 2026B. The pages a brand controls are where a commercial answer is assembled, and those pages carry the product facts the model quotes and describes. When those facts are complete, current, and consistent across a brand's properties, the model has an accurate story to name. When they conflict or are missing, the model has less to be confident about and names a competitor with a clearer record. Structured product answers are the input a brand controls in a process that has moved off its homepage and into the answer.