How AI answers choose which B2B tools to name

A buyer on an engineering team types a question into ChatGPT before they talk to anyone: which tools should we look at for reviewing AI-written code? The answer names three products, says a sentence about each, and links a few pages. If yours is not one of them, you were not on the shortlist, and nobody on your team saw it happen.

Heads of marketing are asking the same thing this year in different words: are we in the answer, and who is instead? This note is about where those names come from, why a single answer tells you very little, how to measure it honestly, and what is worth doing about a gap. It does not promise a way to rank. Nobody can.

Two ways an answer is made

An AI answer about your category is built from one of two places, and often both.

What the model already knows. A model is trained on a large body of public writing up to a cutoff date. If your company, your category and the phrase buyers use for it appeared together often enough before that date, the model can name you from memory. If you launched after the cutoff, or you describe your category in words nobody else uses, it cannot.

What the app finds when it searches. ChatGPT, Claude and Gemini can all search the web before they answer. When they do, the answer is written from the pages the search returned, and the pages are shown as sources. OpenAI and Anthropic both document that their models' answers come back with the sources they drew on attached as citations when they search through the API. The practical consequence: when an app searches, the pages that rank for the buyer's exact question shape the answer more than your homepage does.

What tends to get a tool named

None of what follows is a formula, and we have not measured it as one. It follows from how the answers are made.

• A page that answers the question as it is asked. A buyer asks "what are the alternatives to" a known product, or "which tools fit a team like mine". Pages written to that question, on your site or someone else's, are what a search returns.
• Third-party pages that mention you in the same breath as the category. Review sites, community threads and comparison articles are often what a search finds for a buying question. A tool that appears in them is easier for an answer to name.
• Consistent words for what you are. If your site, your docs and the pages that write about you all use the category phrase your buyers use, both memory and search connect you to the question.
• Recency, when the app searches. A search can surface a page published last week; a model's memory cannot.
• A page the crawlers can read. An app that searches sends a crawler to fetch the pages it might cite. A page that robots.txt blocks for that crawler, or that answers it with an error, cannot be cited however good it is. These crawlers do not run JavaScript, so no analytics tag sees them: your own server or CDN logs are the only record of which ones read which pages.

Why one answer proves very little

Ask the same app the same question twice and you can get two different lists. The app also adds its own instructions, memory and settings, and where the person is asking from changes what a search returns. A screenshot of one answer is an anecdote.

Measuring it honestly takes four things:

1. A fixed panel of buyer questions, grouped by intent: questions that name no brand, questions that weigh options, and questions that compare you with a named competitor. Only the first kind tells you whether you are discovered.
2. Repetition over days. One check a day per question per platform, compared seven days against the previous seven, turns a noisy answer into a rate you can trust.
3. Two numbers, not one. Visibility is the share of answers that name you. Share of voice is your share of all the mentions of you and your competitors. Visibility can hold while share of voice falls, when answers start naming more brands.
4. A label on the method. An answer collected through an official API approximates what the consumer app shows; it is not identical. Say which one you measured, and check the gap against the apps by hand.

How the tools that measure it differ

As of September 2026, on each vendor's own public pages:

• Profound queries nine AI platforms, runs every tracked prompt daily, and covers more than 30 languages and more than 150 regions. It says it captures "directly from the browser", and it adds prompt volumes and agent analytics. If AI answer share of voice is the whole program, it is the deepest instrument.
• AirOps shows ChatGPT insights on its Solo plan and insights across seven or more answer engines on Pro, alongside content operations for AI discovery.
• Jasper sells GEO and AI optimization that monitors citation rates and content gaps, beside its writing platform.
• Viewfield, ours, checks ChatGPT every day on every plan and Claude and Gemini on Team, through each platform's official API with a pinned model, and labels every result by method. It drafts ten buyer questions from your site at setup, and it reads the answers beside your competitors, your buyers and your positioning, so a real change arrives in the morning Brief with why it matters. It also reads your own site the way the answers use it: every page with the prompts it is cited on, its citation share and rank, a readability check, and the AI crawlers that read it, from your own logs. For each prompt it suggests the next step, a change to the page you have or the post that is missing, and saves it to your CMS as a draft for a manager to publish. It covers fewer platforms than Profound, on purpose.

What to do about a gap

Say the answers to a discovery question name two competitors and not you, day after day. That is a content argument with evidence attached: the question, the answers, and the pages they cite.

• Read the cited pages first. They are what the answers are built from. Often it is one comparison article, one community thread, or a competitor's guide that answers the question directly.
• Write the page your buyers are asking for, in your own frame. Not a rebuttal of the competitor, and not a list of keywords: the clearest answer to the question, with the evidence only you have.
• Keep your category words consistent on your site and in what you say elsewhere, so memory and search both connect you to the question.
• Check that the answer engines can read the page. Look at what your robots.txt allows each AI crawler, and at your logs for the crawlers that fetched the page and the errors they were served. A 404 or a 403 an answer engine hit is a page it could not cite.
• If the answers cite a page of yours that does not say who you are, fix that page first. The answers already use the page; it has to say plainly what you are and who it is for.
• Measure it the same way afterwards. Compare the prompts the change targeted with your other prompts over the same weeks, two and four weeks after it went live. A change in the answers after you publish is an association, not proof that your page caused it.

What not to do: chase a single answer, rewrite your positioning around one screenshot, or buy a promise of a ranking.

If you want to see which questions name you and which name your competitors, which of your pages the answers cite, and the change that would put you in them, on your own market, the trial takes fourteen days and no card: start here.