Which companies does AI name for B2B buyers? The B2B AI Shortlist Index

Where the Index stands today.

The B2B AI Shortlist Index measures which suppliers AI engines name when a B2B buyer asks for a shortlist. It covers eight categories of B2B services, 20 buyer questions in each, six engines and three runs of every question on every engine, every month.

There are no results on this page, and there will be none until the first release. We are publishing the method first for a simple reason: an index is only as trustworthy as its method, and a method written after the numbers are in can be bent to fit them. Everything below is fixed for version 1. If we change it, the change is dated in the release notes.

Each release will be free to read, with the raw counts downloadable as a CSV under a Creative Commons Attribution licence, so other people can check our arithmetic and reuse the data with credit.

Why a B2B shortlist index.

Buyers now start supplier searches in AI chat. In a G2 survey of more than 1,000 B2B software buyers published on 6 October 2025, half said they begin their research in an AI chatbot rather than Google, and AI chat ranked first among the sources they use to build a shortlist. Forrester’s State of Business Buying 2026 (21 January 2026) reports that 94% of buyers use AI during a purchase.

What those engines say is hard to pin down. Research by SparkToro with Gumshoe.ai, reported by Search Engine Journal on 30 January 2026, found less than a one-in-100 chance that ChatGPT or Google’s AI would return the same list of brands on repeated runs of one prompt, and concluded that how often a brand appears across many runs is more stable than its position in any one answer. A single screenshot of a single answer proves very little. A fixed panel, run repeatedly, every month, shows a pattern.

The published indexes we know of focus elsewhere. Semrush’s AI Visibility Index (expanded edition, 26 June 2026) covers 126 million prompts and is filtered by industry and platform, built for brands broadly. Studies such as Ahrefs’ analysis of 750 recommendation prompts (4 December 2025), which found “best X” list articles were 43.8% of the page types ChatGPT cited, tell us how engines pick sources, but not which B2B service firms they name in a given month. That is the gap the Index fills: B2B services, buyer-shaped questions, repeated runs, and the sources behind every answer.

The eight categories.

We start with eight categories of B2B services that buyers commonly research with AI. Each has its own panel of 20 questions. The examples below come from the draft panels; the final 20 questions per category are published with the first release and stay fixed after that, so month-on-month movement means something.

Prolinkage offers services in several of these categories. Chapter 08 explains how we handle that conflict of interest.

Questions, engines and runs.

The questions

Each category panel has 20 questions written the way buyers ask them: a mix of plain category questions (“which agencies do X?”), questions with a constraint (company size, sector, budget, country) and problem-first questions where the buyer describes a pain rather than a service. Questions never name a firm. They are written in English and asked as a new conversation each time, with no earlier messages.

The engines

Six engines: ChatGPT with search on, Perplexity, Gemini, Google AI Overviews, Google AI Mode and Microsoft Copilot. Google states that AI Overviews and AI Mode “may use different models and techniques, so the set of responses and links they show will vary”, so we treat them as separate engines.

How answers are collected

Where an engine offers an official programming interface with web search, we use it with fixed settings and record the model name it reports. Where an answer is only available in the consumer product, we use a documented browser method: logged out where possible, a fresh browser profile with no history, memory off, and the location set to the release market. Each release states which method was used for each engine. The first release is run in English from the United States; any added market is announced before it is added.

The runs

Every question runs three times on every engine within one seven-day collection window before each release. That gives 360 answers per category per month (20 questions × 6 engines × 3 runs), 60 per engine, and 2,880 answers across the eight categories.

What we record for every answer.

Matching names to firms is the step most open to error, so it is rule-based and checked by a person. Each firm has a published list of the name variants we accept (for example, with and without “Ltd” or “Agency”). Names we cannot match with confidence go to a review queue rather than being guessed. Every number is computed by a script from the stored records, never typed in by hand.

How share of runs is computed.

The headline measure is share of runs: the proportion of answers in a category that name a firm at least once.

Share of runs (firm, category, month) = answers that name the firm ÷ all answers in the category that month. With a full collection, the denominator is 360. If an answer fails or comes back empty, it is logged, excluded, and the denominator shown with the release is the number of answers actually collected.

Alongside it, every release shows the same measure per engine (out of 60 answers each), the number of engines that named the firm in at least one run, and the raw counts behind every percentage. A firm is listed in the category table once it is named in at least five answers in that month; every name, including those below the line, is in the CSV.

Why share of runs, not rank

Because answers vary so much from run to run, a firm’s position in a single list carries little information. The SparkToro and Gumshoe.ai research reached the same conclusion: visibility across many runs is the steadier signal. We record the order of mention, but we do not publish a ranked “top ten”, and the Index is not a quality rating.

Sources and agreement

For each category and engine we also publish the mix of cited source types (lists, directories, company sites and so on) and the average number of sources per answer. Engine agreement shows how many firms were named by at least four of the six engines. Month-on-month changes appear from the second release, and only where the panel is unchanged.

What the Index cannot tell you.

We would rather state the limits than have a reader discover them.

• It is a sample. 20 questions per category cannot cover every way a buyer asks. Different wording can produce different firms.
• Answers vary. Three runs reduce noise but do not remove it. Small differences between firms in one month should not be read as meaningful.
• Personalisation is excluded. Real buyers are often logged in, with memory and location turned on. We deliberately measure a clean session, which may differ from what a particular buyer sees.
• Interfaces differ. An answer collected through a programming interface can differ from the consumer product, and engines change their models without notice. We record the model where it is shown.
• One market at a time. The first release covers English-language questions from the United States.
• Named is not recommended. Version 1 records whether a firm is named, not whether the answer speaks well of it.
• Correlation, not cause. The Index shows what engines said, and which sources they cited. It cannot prove why.

Fairness, conflicts and corrections.

Prolinkage is an agency, and some of the categories in the Index are categories we work in. We handle that conflict in four ways. The panel of questions is published with the first release and fixed before any collection. Every number is produced by a script from stored answers, and no figure is adjusted by hand, for any firm, ourselves included. The raw counts are published so anyone can recompute them. And firms appear in the Index only as measured facts: “named in this share of answers”, never as a recommendation or a ranking of quality.

No firm can pay to be included, removed or described differently. If you believe a name was matched wrongly, or an answer was misread, write to contact@prolinkage.com with the release month and category. We check every report against the stored answer, correct genuine errors in a dated changelog, and never remove a correctly measured result on request.

Release calendar.

The Index is released on the first Tuesday of each month. Collection runs in the seven days before each release.

1. Release 1Tuesday 5 January 2027
2. Release 2Tuesday 2 February 2027
3. Release 3Tuesday 2 March 2027
4. Release 4Tuesday 6 April 2027
5. Release 5Tuesday 4 May 2027
6. Release 6Tuesday 1 June 2027

What each release will contain

• For each of the eight categories: the firms named in at least five answers, their share of runs overall and per engine, and how many engines named each one.
• The mix of source types the engines cited for each category, and the most-cited domains.
• Per engine: the average number of sources per answer and how often it searched the web.
• From February: what moved since the previous month, where the panel is unchanged.
• The full question panel, the method version, a changelog, and the CSV of raw counts under a Creative Commons Attribution 4.0 licence.

Which category is your company in?

Type your company and website, and pick the category your buyers would put you in. We’ll show you the kind of questions the Index will ask about it, and you can have us run where you sit today in the free AI Visibility Snapshot, without waiting for January. Nothing you type is sent anywhere until you choose to request the Snapshot.

Questions about the Index.

When is the first B2B AI Shortlist Index released?

On Tuesday 5 January 2027, covering all eight categories. After that it is released on the first Tuesday of every month. Until then, this page publishes the method only, with no results.

Which AI engines does the Index cover?

Six: ChatGPT with search on, Perplexity, Gemini, Google AI Overviews, Google AI Mode and Microsoft Copilot. Each question runs three times on each engine every month, and every release states how each engine’s answers were collected.

Can a company pay to appear in the Index?

No. Inclusion, removal and descriptions cannot be bought. Firms appear only because an engine named them, and the numbers come from a script run over stored answers. Genuine errors are corrected in a dated changelog.

Is the Index a ranking of the best agencies?

No. It measures how often AI engines name a firm for a fixed set of buyer questions. That reflects what the engines read, not the quality of anyone’s work, and we do not publish a ranked list.

Can I use the data?

Yes. Each release includes a CSV of raw counts under a Creative Commons Attribution 4.0 licence: reuse it freely, with credit to the B2B AI Shortlist Index by Prolinkage and a link to the release.

How do I find out where my own company sits before January?

Request the free AI Visibility Snapshot. We run a sample of your buyers’ questions on the main AI engines and show who gets named, which sources are cited and whether you appear. It is human-reviewed and arrives within 24 hours.

Sources: G2 buyer survey, 6 Oct 2025; Forrester, State of Business Buying 2026, 21 Jan 2026; Search Engine Journal on SparkToro and Gumshoe.ai research, 30 Jan 2026; Semrush AI Visibility Index, 26 Jun 2026; Ahrefs, 4 Dec 2025; Google Search Central, AI features and your website (accessed 7 Nov 2026); Creative Commons Attribution 4.0 licence. Method: B2B AI Shortlist Index, method v1, Prolinkage, 8 Nov 2026.