Guides · Updated October 31, 2026
What is AI visibility, and how to measure it
AI visibility is how often AI assistants like ChatGPT and Google's AI Overviews and AI Mode name your product when people ask the questions that lead to buying it. There's no standard way to measure it: every tool computes its own score, and the same answers can score 40% in one and 50% in another. Here's what's worth counting, how to count it so the number means something, and how it connects to signups and revenue, with our own product's numbers as the example.
What AI visibility means
Two different things go by the name, and it helps to keep them apart:
- Product visibility: does the answer name your product, and how? This is what a founder means by “does ChatGPT recommend us?”
- Page visibility: is one of your pages a source the answer cites? This is the only thing Google and Bing report to site owners.
They come apart more than you'd think: in one controlled test, 43% of answers that cited a page about a new brand never named the brand (Ahrefs). The terms you'll meet:
- Mentioned: the answer names your product at least once, anywhere.
- Recommended, or top pick: the answer offers it as an option, or as its main choice. Most headline numbers count a passing mention the same as a recommendation; look for the two counted apart.
- Cited: one of your pages is among the answer's sources.
- Share of voice: your mentions as a share of all products' mentions.
- Position: where you come in the answer's list.
Same answers, different scores
Take 10 answers: 4 name you, and 2 name no product at all.
- Peec AI divides by all answers: 40% (Peec AI's docs).
- Profound leaves out answers that name no product: 50% (Profound's docs).
- Semrush and HubSpot turn it into a 0–100 score from several parts, without publishing how they're weighted.
So scores can't be compared across tools. Two other things to watch for. Average position is usually averaged only over the answers that name you: in Peec's own example, a product named in 6 of 10 answers, first in 3, has “position 1.5” while it was first in only 3 of 10 (Peec AI). And a single 0–100 score hides which answers it came from, so you can't check it.
A worked example: PodSeeker
PodSeeker, a podcast database for PR teams, is our own product. ChatGPT's answers to its 25 buying questions in the week to October 29, 2026 (102 answers, about 4 per question):
| Product | Named | Top pick | Its pages cited |
|---|---|---|---|
| PodSeeker | 18 of 25 questions | 14 of 25 | 21 of 25 |
| Rephonic | 12 of 25 | 0 | 8 of 25 |
| Podchaser | 12 of 25 | 0 | 7 of 25 |
| Listen Notes | 10 of 25 | 0 | 3 of 25 |
A question counts when most of its answers that week do. What the table shows:
- Cited isn't named. PodSeeker's pages were cited on more questions (21) than it was named on (18): on some, ChatGPT read our pages and named someone else.
- Named isn't recommended. It was the top pick on 14 of the 18.
- Questions, not answers. Counted per answer, PodSeeker was named in 80 of 102 (78%). Counting questions gives every question the same weight however often it's asked, and keeps each number checkable: you can open the answers behind any question.
- Who wins instead matters as much as your own number: the 7 questions PodSeeker missed are the ones to read, to see who ChatGPT named and which pages it read.
- Each assistant is its own number. We started asking the same 25 questions on Google's AI Overviews and AI Mode on October 31. On that first day, one answer each and far too early to call, AI Overviews named PodSeeker on 15 and AI Mode on 7, against ChatGPT's 19 that week. Added together, those would hide where to work (why they differ).
How to measure it well
- Ask the questions your buyers ask. Unbranded buying questions with the buyer's situation in them (“best podcast database for a small PR agency”), not keywords. Wording matters more than anything else: rewording a question changes which products come back more than asking it again does (arXiv 2605.27440). Track branded questions (“PodSeeker vs Podchaser”) separately; they name you almost every time.
- Freeze the list. A number only means something over time if the questions stay the same. When you change one, note the date.
- Ask each question several times. ChatGPT gave the same list of brands less than 1 time in 100 across thousands of runs, but how often each brand appeared held steady (SparkToro). Three or more answers per question per week, decided in advance.
- Count presence, not rank. Whether you're named holds up across runs; your position churns. Show position as a count over all answers (“first in 3 of 10”), never as an average over the answers that name you.
- Hold the conditions fixed. Ask where buyers ask: the ChatGPT app (the app and OpenAI's API share only a small share of the pages they cite: arXiv 2609.18729), and Google's own search page for AI Overviews and AI Mode. Logged out, so your own history doesn't shape the answer, from one country and language, and note the model and date of every answer. A model change alone moved one brand from 39% to 99% of answers (Citation Labs).
- Count each assistant apart. ChatGPT, AI Overviews and AI Mode often recommend different products for the same question, so a blended number moves when one of them does and tells you neither.
- Use many questions. Pinning one question's rate to ±5 points takes about 385 answers; spreading answers across 15–30 questions gets most of that precision cheaply (Gumshoe).
With your agent, in Askadia (ours)
Give Claude Code, Cursor, Codex or any agent one line, Read https://askadia.ai/docs/api.md, and approve it once, or
add Askadia's MCP server (in Claude Code or Cursor, or as a connector in Claude.ai or ChatGPT). Then ask it to:
- Review your questions against the list above: a real buyer, neutral wording, specific, the way buyers ask. It says which need work and suggests better ones; you decide.
- Add, pause and resume questions, and set each one's assistants and how often on each: daily for a baseline, Mon/Wed/Fri once it's steady, daily again for a launch.
- Read every answer word for word on each assistant, with the products named in order and the pages cited, and tell you what changed since last week.
Included in the $29 plan, and reading is unlimited. It proposes changes until you tell it it can act, and the plan and its answers stay your click.
Real change or noise?
On PodSeeker's 24 questions with 3 or 4 answers that week, ChatGPT named it in every answer to 16 and in none to 2; on the other 6 it changed its mind. Questions like those flip from week to week by chance. Simple arithmetic shows how often a question's majority verdict changes between two weeks when nothing has really changed:
| Named in, truly | Flips by chance, 3 answers a week | 7 answers a week |
|---|---|---|
| 50% of answers | 50% of weeks | 50% |
| 70% | 34% | 22% |
| 90% | 5% | under 1% |
Our own calculation, assuming independent answers (real ones are somewhat correlated). Before you believe a change:
- Count how many questions moved, and in which direction: chance flips go both ways and cancel out.
- Wait for it to hold two weeks or more.
- Ignore moves of one question on a question that was already split.
- Compare with your baseline, not just last week.
- Check what else changed: your questions, the model, the app.
What Google and Bing show you
The engines' own reports are free and worth setting up, but they cover your pages, not whether your product is recommended:
- Google Search Console shows how often your pages appear in AI Overviews and AI Mode, as impressions only (Google).
- Bing Webmaster Tools shows citations of your pages in Copilot and Bing's AI answers, and the searches behind them (Bing).
- OpenAI offers no dashboard; ChatGPT's links carry
utm_source=chatgpt.com, most of the time.
None of them tell you, when a buyer asks which tool to buy, whether yours is named, and who is named instead. That's what asking the questions yourself is for.
From visibility to signups and revenue
Clicks from AI assistants are still a small share of traffic, about 0.3% of web traffic from ChatGPT in mid-2026 (SE Ranking, 101,574 sites), so a site with 10,000 visits a month sees about 30. Most of the effect arrives some other way: people shown a brand in ChatGPT were 2.5 times more likely to visit it within a week, but only about 9% of those visits came from ChatGPT itself; most came through a Google search or by typing the address (Similarweb, finance, travel and beauty, not software). At n8n, analytics credited about 1% of conversions to AI search while “how did you hear about us?” credited about 9%.
So track it in more than one place:
- AI-referred visits: Google Analytics has an “AI Assistant” channel since May 2026 (Google).
- Source on the customer: save each signup's first referrer and UTM on the account, and in Stripe's customer metadata, so revenue has a source.
- “How did you hear about us?” with ChatGPT, Claude, Perplexity and Gemini as separate options, and a box for what they asked.
- Searches for your name in Search Console, the “heard it in ChatGPT, then Googled it” path.
- Notes on the breaks: ChatGPT dropped its link tag for about a week in October 2026, and reported ChatGPT visits fell 30–40% while the visits went on (Profound).
Then put your questions count next to these each week, and look for them moving together over months.
What nobody knows yet
No study has tied a product's mention rate to its revenue, isolated how much a logged-in user's memory changes the answer, or followed a small software product from unrecommended to recommended. And almost every number in this field comes from a company selling the measurement, us included: we sell repeated question tracking, and this guide concludes you need it. Ask any tool, ours too, for its method, its sample and the raw answers. Ours is here.
Related: How to rank in ChatGPT · Best AI visibility tools · Why ChatGPT doesn't recommend your product
Measure yours, free
The free checkup asks ChatGPT, Google AI Overviews and AI Mode 10 questions your buyers ask, in about 2 minutes, with every answer in full.
100% free: no email, no sign-up, no credit card · Takes about 2 minutes