Research · 38 pages
From answer to checkout: what the evidence says about getting a B2B company recommended by AI assistants
A graded review of the 2023–2026 evidence on how AI assistants choose which B2B companies to recommend and how a recommendation becomes a purchase. It separates what replicated from what did not, and ends with a program graded step by step.
Key findings
- 01
AI assistants cite pages they retrieve from a search index, so a page that Google, Bing, or an assistant's own crawler has not indexed cannot be cited, however well it is written (strong: platform documentation).
- 02
Once pages are retrieved, relevance to the question and position among the sources drive citation in controlled tests, and stated prices, real dates, and an answer near the top of the page help (moderate: laboratory studies).
- 03
Most rewriting tactics did not replicate: the original GEO study's gain of up to 40% held for a page already in a simulated engine's context, and a later benchmark found significant gains in only 3 of 54 cases (strong: peer-reviewed benchmarks).
- 04
Assistants cite third-party pages far more often than a company's own, and mentions of a brand on other websites and on YouTube tracked its AI visibility far more closely than backlinks did in a study of 75,000 brands (moderate: correlational vendor data and preprints).
- 05
AI recommendations behave like distributions, not rankings: the same prompt rarely returns the same list twice, so visibility has to be measured as a rate over repeated runs on each assistant (strong: vendor, preprint, and peer-reviewed studies agree).
- 06
B2B buyers use assistants to build shortlists and then check what they were told, most often on the vendor's own website (moderate: vendor surveys and one usability study).
- 07
As of September 2026, no major assistant documents a way to sell a professional service inside the chat, so the purchase happens on the seller's own checkout, and in the documented flows that use a buyer's card a person approves the payment (strong: platform documentation).
- 08
Google's spam policies cover attempts to manipulate AI answers, Bing may remove such content from its search experiences, and the FTC's rule bans fake reviews, including AI-generated ones, and incentives tied to a review's sentiment (strong: primary policy and regulatory texts).
- 09
No study found for this review measures how long a new company takes to be recommended, and a 2026 review of 45 studies found no technique with a lasting causal effect on discoverability across engines (an evidence gap, not proof of no effect).
Preview
Author: Paul Maxwell, founder of AEO HQ. Version 1.0, September 2026. Every number and quotation in this paper was checked against its source on September 27 and 28, 2026.
Executive summary
Many B2B buyers now begin vendor research in an AI assistant. This paper reviews the evidence published between September 2023 and September 2026 on two linked questions: what makes an assistant name a company, and what has to happen between that answer and a purchase. It grades the evidence behind each finding and keeps findings separate from AEO HQ's recommendations.
The evidence supports a short chain.
- Retrieval is the gate. ChatGPT, Google's AI features, Copilot, Claude, and Perplexity each document that the pages they cite come from a search index, their own or a partner's. A page that is missing from the relevant index cannot be cited through search.
- Among retrieved pages, relevance and position matter most. In controlled tests, relevance to the question and position among the retrieved sources drove citation, and specific facts such as prices and dates helped. Most rewriting tactics sold as generative engine optimization (GEO) did not replicate. The best-known result, that GEO "can boost visibility by up to 40%", was measured for a page already placed in a simulated engine's context, and a later benchmark found statistically significant gains in only 3 of 54 cases.
- Other sites carry the recommendation. Assistants cite third-party pages far more often than a company's own, and in the largest study, mentions of a brand on other websites and on YouTube tracked its visibility in AI answers far more closely than backlinks did. This evidence is correlational.
- Recommendations are distributions. The same prompt rarely returns the same list twice, so AI visibility has to be measured as a rate over many runs, not read from one answer.
- Buyers verify, then buy elsewhere. B2B buyers use assistants to build shortlists and then check what they were told, most often on the vendor's own website. No major assistant documents a way to sell a professional service inside the chat, so the purchase happens on the seller's own checkout.
Two sets of limits frame all of this. Google's spam policies cover attempts to manipulate AI answers, and Bing warns that such content may be removed from its search experiences. The U.S. Federal Trade Commission's rule on fake reviews covers AI-generated reviews and incentives tied to a review's sentiment.
The program that follows from the evidence is ordinary and specific: be indexed, state the offer as checkable facts, keep those facts consistent everywhere, earn independent corroboration, make the checkout completable without a sales call, measure with repeated runs, and stay inside the rules. Section 4 grades each step. No study has measured how long any of this takes, and this paper makes no promise about timing.
AEO HQ sells answer engine optimization (AEO) audits and implementation, so it has an interest in these conclusions. Section 7 describes that interest.
1. Introduction
1.1 Why the question matters
Three recent surveys show the scale of AI use in B2B buying. In a March 2026 survey of 1,076 B2B software buyers, 51% said they start their research with an AI chatbot more often than with Google, up from 29% eleven months earlier (vendor survey). Among 519 U.S. B2B professionals who use AI at work, agencies and service providers were the category most often researched with AI, at 51% (vendor survey, March–April 2026). Gartner found that 45% of 646 B2B buyers had used AI during a recent purchase (August–September 2025). The figures differ because the samples and definitions differ, as section 3.5 shows. All three show substantial use.
For a B2B company, and especially one that sells services, two practical questions follow: what makes an assistant name the company when a buyer asks for options, and what has to be true, once the company is named, for the buyer, or an agent acting for the buyer, to complete a purchase. Advice on both questions is plentiful and mostly untested. Google lists services "Promising improvements for AI experiences and search formats (also known as 'AEO' or 'GEO' tools)" among those site owners should evaluate carefully, and it says third-party tools "can't guarantee performance". This paper sets out what the evidence supports, what it contradicts, and where it is silent.
1.2 Definitions
- Answer engine optimization and generative engine optimization. Two names for overlapping work aimed at being cited, described accurately, and recommended in answers written by AI systems. Google calls AEO and GEO "both terms you may see used to describe work specifically focused on improving visibility in AI search experiences". This paper uses AEO for the practice and GEO where the research literature uses that name.
- AI assistant. ChatGPT, Google's AI Overviews and AI Mode, Gemini, Claude, Perplexity, and Microsoft Copilot.
- Parametric knowledge and retrieval. Parametric knowledge is what a model learned in training. Retrieval is the step in which an assistant searches an index and reads pages before it writes. Basing an answer on those pages is called grounding.
- Query fan-out. Turning one question into several searches. Google describes AI Overviews and AI Mode as "issuing multiple related searches across subtopics and data sources".
- Mention, citation, and recommendation. A mention is a company's name in an answer. A citation is a link to one of its pages shown with the answer. A recommendation is a mention offered as an option for the buyer's need.
- Agentic commerce. Buying in which an AI agent carries out some of the shopping or checkout steps for a person.
1.3 Scope
The review covers unpaid visibility in the six assistants above, not advertising inside them. Its focus is B2B companies, with particular attention to firms that sell services at a fixed price, because that is the case AEO HQ is built around and the case the research covers least. Legal points refer to the United States. The diagram below shows the path the paper follows.
From a buyer's question to a paid order
Assistant
- 01Decide whether to searchSearch does not run for every question
- 02Retrieve from an indexPages missing from the index cannot be cited
- 03Rerank and selectRelevance and position weigh most in tests
- 04Write and citeAnswers vary from run to run
Buyer
- 01Read the answerNames vendors, sometimes with links
- 02Verify the claimsMost often on the vendor's own website
- 03Check other sourcesSearch results, reviews, and comparisons
- 04Buy on the seller's siteNo assistant documents in-chat service sales
- ChatGPT queries with web search, February 2026
- 34.5%
- AI-using B2B buyers who visit a vendor's site after an AI mentions it
- 71%