What AI assistants say about your product, and how to audit it

Semrush asked 2,338 US adults about shopping with chatbots in September 2026. Among people who use an assistant when they shop, 81% had decided against a purchase because of what it told them, and 74% of all respondents said mixed or negative reviews surfaced by a chatbot would make them less likely to buy. That is consumer data, but the mechanism is the same in B2B and the stakes are higher: a buying committee member asks an assistant to compare three vendors, gets a summary of your G2 reviews, a Reddit thread from 2024 and your competitor’s comparison page, and you never see the session because it ended without a click.

Most GEO work still optimises for one thing, being cited. This note is about the other half of the problem, which is what the assistant says once it has you in the answer, and how to run a standing audit of that rather than a one-off screenshot exercise.

Why citation counts miss the problem

Comscore’s Q2 2026 report separated two numbers that vendors usually blur. Sourcing is how often an assistant reads a domain while building an answer. Citation is how often it names that domain. For travel, Tripadvisor accounted for 61% of sourcing and 21% of citations. An assistant can retrieve your page every time, fold your pricing into the comparison, and attribute nothing to you. A citation-rate dashboard shows a flat line while the assistant repeats a two-year-old complaint about your onboarding to every prospect.

The second thing citation counts miss is the split across assistants. Comscore has ChatGPT’s share of prompt volume falling from 70% to 50% between January and June 2026, with Gemini at 30% and Claude at 11%. Gemini sits inside Workspace, Claude skews to knowledge workers, and a B2B buying committee is spread across all three. An audit that only queries ChatGPT covers half of the conversations your buyers are having.

If you are still deciding which visibility vendor to trust, AI visibility metrics: diagnostic vs decision-grade covers what makes a number fit for a forecast. This note assumes you want the raw answers, weekly, in a table you own.

What to ask, and how often

Build a prompt set of 20 to 30 questions in three groups, written the way a buyer types, with no brand-favourable phrasing.

The first group is category questions: “best [category] for a 200-person B2B SaaS company”, “alternatives to [market leader]”, “[category] tools that integrate with HubSpot”. These tell you whether you appear at all and which competitors the assistant reaches for by default.

The second group is direct questions about you: “is [product] worth it”, “[product] pricing”, “[product] vs [competitor]”, “problems with [product]”, “[product] reviews”. These are the ones that produce the sentence that loses the deal. Include the negative framings on purpose, because buyers use them.

The third group is objection questions built from your win-loss notes: “does [product] work with Salesforce”, “[product] data residency EU”, “how long does [product] take to implement”. If your sales team hears it on a first call, the assistant is being asked it a week earlier.

Run each prompt against ChatGPT, Gemini and Claude, through their APIs with web search or retrieval enabled, since the consumer apps and the raw models answer differently. Weekly is the right cadence. Answers move when a competitor ships a comparison page, when a review batch lands on G2, or when a Reddit thread gets upvoted, and none of those wait for your quarterly review.

The Clay build

One table, one row per prompt per assistant per week. The columns that matter:

The full answer text, stored, because you will want to diff it later. A boolean for whether you were named. The list of competitors named, extracted by a second Claude call with a fixed schema. A sentiment score on any sentence about you, on a three-point scale, from the same extraction call. A factual-accuracy flag: the extraction call is given your current product facts (pricing tiers, integrations, security certifications, implementation time) and asked to mark any claim in the answer that contradicts them. And the sources the assistant cited, as URLs, so you know which page to fix.

The extraction prompt is the piece worth getting right. Give it your fact sheet as structured data, ask it to return only claims that contradict the sheet with the exact quoted sentence, and forbid it from inferring anything the answer did not say. Score every row against the same sheet so the number is comparable week to week.

In n8n, the same shape is a scheduled workflow: iterate the prompt set, call each assistant, pass the answer to the extraction step, write to a Google Sheet or back into Clay, and post a Slack digest on Monday with three lines: prompts where you were not named, prompts with a factual error, and prompts where sentiment dropped since last week. The Slack message is the deliverable your team reads; the table exists so you can trace a change back to the week it happened.

Cost is small. Thirty prompts across three assistants with an extraction call each is roughly 180 model calls a week, well under £20 a month at current API prices. Put a token budget on it anyway; the reasons are in the agent cost note.

Fixing what the audit finds

Each finding maps to a source page, and the source page is what you fix, because the assistant will keep repeating whatever it retrieves until the retrieval changes.

A factual error almost always traces to a stale page, and it is often yours: a pricing page cached from before the last change, a docs page that still lists an integration you deprecated, a partner listing nobody has touched in two years. Correct the page, then ask the assistant again in a fortnight. If the error comes from a third-party page, a review site or a comparison post, the fix is to publish the accurate version somewhere the assistant retrieves with higher weight, usually a page on your own domain with the specific question as the H2 and the answer in the first sentence.

A negative sentiment line usually traces to reviews. The answer to “problems with [product]” is a paraphrase of your three-star G2 reviews, and the assistant does not weight them by date. The fix is volume and recency: a structured review campaign with customers who have used the feature the complaints mention, so the next retrieval sees twenty recent four-star reviews that describe the fix rather than three old ones that describe the problem. Reddit works the same way, and a founder answering a two-year-old thread with what changed is retrieved.

A competitor appearing where you do not is a content gap, and the assistant tells you the exact shape of it: the competitor has a page that answers the category question and you do not. The page you write needs the question in the title, a direct answer up top, and named entities the assistant can match, the tools you integrate with, the company sizes you serve, the certifications you hold. Citation infrastructure covers the page structure that gets retrieved.

What the number is for

The audit gives you two metrics for the board. The first is accuracy: the share of direct-question answers with zero contradictions against your fact sheet, per assistant. Start from whatever it is, and expect it to be below 70% the first week. The second is the share of category answers that name you, which is the closest thing to share of voice you can measure without a vendor.

Neither number is attribution. Whether any of this moved pipeline is the separate question covered in the attribution gap nobody’s measuring, and the honest answer for the first two quarters is that you will see it in branded search lift and in the form field before you see it anywhere else. The audit’s job is narrower: make sure that when a buyer asks an assistant about you, the answer is true, current and yours.