How to measure AI answer share

A metric that uses a new question every week cannot rise or fall. It can only change subject. Write the prompt set down: the category questions, the alternatives questions, the fit questions, the constraint questions. Give each an ID, a buyer, a country, and a language. Run that list. When you add a prompt, mark it as new so it does not pretend to be history. When you retire one, keep the old rows so a chart does not silently drop your worst or best question.

Wording is part of the freeze. "Best billing software" and "billing software for a usage-based SaaS with usage invoices" are different markets. Small wording edits are allowed when you are correcting a typo. They are not allowed as a way to hunt a flattering answer. If you want to know sensitivity, run a variant as an extra row. Do not replace the original. The prompt design notes explain how to choose the list from deals rather than from a brainstorm.

Sample size is the number of prompts that match pipeline, not the largest number you can afford to paste. Thirty serious prompts beat three hundred generic ones because you can read the answers. If you cannot read the description on the prompts that moved, you do not have a measurement system. You have a scraper.

Ask the same assistant the same question twice and you can get two different shortlists. Answers are sampled, retrieval shifts, and the order of names wobbles. One run per prompt per week measures luck as much as position. So ask each prompt several times on each engine, five is a sensible default, and count answers, not prompts.

Then report a range, not a point. With 30 prompts asked 5 times on one engine, 150 answers, a 40 percent mention rate is honestly somewhere around 32 to 48 percent, and wider still if a few prompts drive most of the mentions. A move from 40 to 44 inside that range is noise. The range narrows as you add asks, which makes it the honest way to show how sure you are.

Treat a move as real only when the change clears its range and is big enough to matter. Compare against the last few weeks rather than last week alone, so one odd week cannot fake a trend. When an engine updates its model, expect many prompts to move at once, and mark that week, because the change is the engine, not your work. The GEO agent does this math for you: every number comes with a likely range, position gets one once at least three answers name you, and alerts fire only on moves beyond the usual swing, with a flag when the engine itself changed.

Mention: yes or no, your brand, in this answer. Mention rate is the share of answers that mention you on that engine that week, counted across every ask of every prompt. It is the headline and the easiest number to game, which is why it never stands alone. A mention in a warning is still a mention. Read it.

Position: where the name sits when the answer ranks options. First named, inside a short list, or only in a source the reader must open. Never average "first" and "absent" into a meaningless middle. Average position only over the answers that name you, show it beside the mention rate, and give it a range, because a position built from two answers is a guess.

Accuracy: does the paragraph match the facts block. Right category, right buyer, right limits, no invented features. Score it pass or fail with a note, and sort the fails by type, since a wrong fact, an outdated one, and a lookalike company's details each need a different fix. A rising mention rate with falling accuracy is a bad week. Sentiment is the softer cousin: are you recommended, described neutrally, or warned against. Keep sentiment, and do not let it replace the fact check. A warm paragraph that quotes the wrong price is a defect.

Cited sources: every URL the engine showed for that prompt. This is the operational field. It tells you what to pitch, what to correct, and whether your own page has entered the retrieval set. Store the URL, the domain, the type of source, and whether you are the subject of that page. Over a month, a domain frequency table is more useful than any single answer.

Two more numbers fall out of those fields. Share of voice is your share of all brand mentions across the answers, which shows a competitor crowding you out even when your own rate holds. And the list of brands named instead of you, with the sources cited when they were, is the most useful competitive research you will read all month.

ChatGPT, Perplexity, Gemini, and Google AI Overviews do not move together. Live-search products can reflect a new cited page within weeks. Answers that lean on accumulated memory move on a longer cycle. A single "AI share" number will rise because Perplexity moved and will convince you the whole market moved. Then a buyer using a different assistant will describe a shortlist you do not recognize.

Start with those four, because they cover how most buyers and shoppers ask today, and add Claude, Grok, Google AI Mode, or Copilot when your leads name them. Copilot has no public API, so any tool that reports it is using a stand-in, such as answers grounded in Bing Search. The label should say so; in the GEO agent it reads "Copilot (Bing proxy)". Do not add an engine because a tool supports it. Each engine costs reading time and money, which is why the GEO agent shows spend per engine against a cap. The weekly review should fit in a meeting a founder will attend.

Country and language are part of the engine breakout when you sell in more than one market. An English-language US prompt set will flatter a company that only wins US answers and hides a broken localized page. Track the markets on the brief. The technical guide covers why the indexed URL for that market has to match.

Run the set on the same weekday. Models and search results jitter. A fixed schedule plus a fixed list makes jitter visible as noise around a trend, instead of as a story you invent midweek. Keep the raw answer, and record the model version, so an engine update does not pass for progress. You will need the answer when someone asks what the model "meant." Summaries without the paragraph are how accuracy bugs survive.

The weekly note has five lines. Mention rate by engine, with its range, versus the last few weeks and the baseline. Prompts that changed position. Accuracy fails. Cited domains that are new or gone. What shipped in the last seven days that could explain a move. If nothing shipped, say so. Correlation is not a trophy, and it is still the only way to decide what to do next.

Do not refresh prompts hourly and do not refresh them quarterly. Hourly readings drown the team in variance. Quarterly readings let you pay for placements you should have stopped. Weekly is the compromise the GEO agent is built around: enough resolution to steer, enough calm to edit a page properly.

Answer share is a leading indicator. Pipeline is the reason you funded it. Ask every demo, trial, and inbound call how they heard about you, and give them a way to say an assistant's name. The inbound agent can record "ChatGPT recommendation" or the caller's actual words as the source, along with the transcript. That tag will be incomplete. People forget. It will still be truer than a modeled multiplier. Referrer data is the other half: the GEO agent's tracking snippet notes visits that arrive from ChatGPT, Perplexity, Gemini, and other assistants, and a signed webhook from your forms credits the lead, with a CSV export for HubSpot or Salesforce.

Watch branded search beside the tags. A lift in people searching your name, especially with pricing and versus modifiers, often follows a lift in mentions. It is not proof of causation. It is a second witness. If mentions rise and branded search and tagged leads stay flat for a month, read the descriptions. You may be mentioned as a minor or mismatched option. That is a message problem, or a prompt-list problem, not a reason to buy more placements.

Report a short narrative to anyone who owns budget. These prompts match our deals. These engines moved. These sources were cited. These pages we shipped. These leads named an assistant. Next month we will correct this domain and rewrite this page. A narrative with those nouns beats a gauge widget. If you cannot fill the nouns, you are not ready to scale spend.

The first full run is the baseline. Everything later is a delta against it. Targets should be prompt-level: named in most answers, with the whole range above half, on the five alternatives prompts that match the largest deals; a passing accuracy score on every mention; your own URL or a fair third party cited on the category prompts. A target of "be everywhere" is a mood. A target on a named prompt is a project.

Nobody can guarantee a model's sentence. Publish that limit in the report so a good week does not become a promise you regret, and show every number with its range, so the reader can see how sure you are. What you can guarantee operationally is the rerun, the stored answer, the accuracy review, and a decision on the next asset. That is the professional standard. Vendors who promise the top of ChatGPT are selling a screenshot service.

Illustrative charts are fine in a sales conversation when they are labeled as a model of how live-search engines tend to move earlier than memory-heavy answers. They are not your dashboard. Your dashboard is your prompt IDs. If an agency shows you a smooth curve and will not show the prompts, assume the curve is a drawing.

The flattering prompt. Someone adds "why is [your brand] the best" and reports a mention rate of 100 percent. Exclude prompts that presuppose you. Measure prompts a buyer could ask without already being yours.

The moving window. Old failures drop out of a "last ten runs" widget because the list changed. Keep a cohort. The prompts you started with in the baseline stay visible even when they hurt.

The single sample. One ask per prompt, a ten-point jump, a celebration. Ask again and half the jump is gone. Ask several times, report the range, and celebrate only the moves that clear it.

The unread win. Mention rate up, and nobody opens the paragraph that calls you a feature you deprecated. Require an accuracy pass as a field that can block a green status. The person who checks can be an operator. It cannot be the same automated summary that is being judged.

The channel with no source. Leadership asks what to do on Monday and the report has no URLs. A GEO report without cited sources is a weather comment. The action lives in the URL list: edit, pitch, or ignore.

A spreadsheet is enough at the start. Columns: date, engine, country, prompt ID, ask number, prompt text, mention, position, accuracy, sentiment, cited URLs, notes, shipped change. One row per answer, so several rows per engine per prompt each week. A second tab pivots mention rate. A third tab counts domains. Store raw answers in a folder named by date. You can outgrow this. You should not start with a platform that hides the row.

When the row count hurts, that is the moment for an agent. The GEO agent asks the set on eight engines, several times each, and fills mention, position, sentiment, sources, and accuracy against your facts file, with a second pass that confirms each problem before it reaches you. Every number carries its range, and competitors, alerts, and spend sit on the same dashboard, with a PDF report every Tuesday. The final call on whether a description matches the facts stays with a person who knows the product. Automation that files a wrong description as positive sentiment is worse than a slow spreadsheet.

Review it next to the SEO numbers for the same topics. Rankings and pipeline of the matching pages, plus answer share of the matching prompts, on one page. The shared brief is what makes those rows comparable. If the topics diverge, the measurement will diverge, and you will fund two stories again. Measure the shortlist you are trying to join, weekly, in the engines your buyers named, and keep the paragraph.