We are on G2 but AI still ignores us

What is actually happening

You are looking at the wrong question. A G2 profile answers "is AI Knows Us any good". Your buyer types something closer to "best onboarding software for a 30 person insurance broker in Manchester". That is a category question with no brand in it. The assistant has to build a shortlist from scratch, in a few seconds, from pages it can fetch and read.

Three things decide whether your name survives that process.

1. The machine has to be able to fetch your pages

Assistants use named crawlers and fetchers. They are documented, and they are not all the same bot. Blocking one in robots.txt, or sitting behind a bot filter that returns a challenge page, is a quiet way to be absent.

Last verified: 18 October 2026

Google's own guidance for site owners is that pages shown in AI features follow normal Search eligibility and the usual preview controls, which is set out in AI features and your website. So an indexing problem is also an AI problem.

2. The machine has to find a sentence worth quoting

Models quote specifics. A price with a date and a tax note. A row in a table. A named limit. A count. "Trusted by hundreds of teams" is unquotable. "Flat 2,400 rupees per seat per year, GST extra, as of 18 October 2026" is quotable. Directory profiles are built for human scanning, not for lifting exact claims, and a lot of your real evidence (integration list, data residency, onboarding time, who you are wrong for) usually is not on them at all.

3. Other sites have to say your name in the same breath as the category

Assistants lean on pages that already group options together: comparison posts, association member lists, local roundups, trade press, forum threads, documentation of partners. If nobody else writes your name next to your category words, the model has nowhere to pick you up. One profile on one review site is a single source. We go into the mechanics in #.

How to confirm it is happening to you

Do not ask an assistant "what do you think of [your brand]". The moment you put your name in the prompt, you have told the model the answer. The reading becomes worthless. Use blind questions only.

1. Write 15 buyer questions that never contain your brand name. Use the words a buyer uses: category, size, place, budget, constraint. Example shape: "[category] for a [size] [industry] in [city], under [budget]".
2. Pick the surfaces your buyers actually use. A reasonable set is ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews.
3. Open a fresh session for each question. Turn off memory or personalisation where the app allows it. Logged out, if you can.
4. Ask each question once per surface. Record the full answer text and the cited links, not just a yes or no.
5. Mark a mention only when your business is named in the answer body. A link in a source tray is a separate, weaker thing. Record it separately.
6. Count mentions over total asks. That is your rate.
7. Put a 95 percent range on the rate before you tell anyone the number.
8. Repeat the whole set on a fixed day each month, with the same questions.

A worked example you can repeat

Say you ask 15 questions on one surface. You are named in 4 answers.

• Rate: 4 divided by 15 equals 26.7 percent.
• 95 percent Wilson interval: 11 percent to 52 percent.

Eight weeks later you repeat it and get 9 of 15.

• Rate: 60 percent.
• 95 percent Wilson interval: 36 percent to 80 percent.

It looks like you more than doubled. You did not prove it. 11 to 52 and 36 to 80 overlap, so the honest report is: no measured change. This is why sample size matters more than dashboards. Run 40 questions instead.

• Before: 11 of 40 equals 27.5 percent, range 16 percent to 43 percent.
• After: 24 of 40 equals 60 percent, range 45 percent to 74 percent.

Those ranges do not overlap. Now you can say it moved. The arithmetic is the Wilson score interval, and you can check every number above in any statistics package or by hand.

What to change, in order

Order matters, because the later items do nothing if the earlier ones are broken.

1. Unblock the fetchers. Read your robots.txt and your CDN bot rules against the table above. Fetch your own key pages with a plain request and confirm you get HTML with the content in it, not a challenge page or an empty shell that needs JavaScript.
2. Write one page per buyer question, not per feature. If the blind question is "onboarding software for a 30 person broker", there should be a page whose title is close to that sentence.
3. Put the quotable facts in the page body. Prices with a date and a tax treatment. Counts. Named integrations. Supported countries. Response times you actually meet. A table beats a paragraph.
4. Say who you are wrong for. Models reward pages that set boundaries, and so do buyers. A short "not a fit if" section makes the rest of the page more credible.
5. Get named elsewhere, on the record. Ask for inclusion in category roundups, association directories, partner documentation, local lists and trade coverage. Each one is a separate source the model can reach.
6. Keep the G2 profile accurate anyway. It is still read by humans at the shortlist stage, and the facts on it should not contradict your site.
7. Re-measure on the same questions. Changing the questions between rounds destroys the comparison.

How long before it moves

Be patient in a specific way. Three separate clocks are running: the crawler has to come back, the index has to update, and the assistant has to start choosing your page over the one it used before. You control the first a little, the second barely, the third not at all.

A sensible cadence is a fixed monthly re-run of the same blind question set, and a decision rule set in advance: you act when two readings stop overlapping, not when one number looks nicer. Anything faster than monthly on a 15 to 40 question set is noise dressed up as a trend.

Where AI Knows Us fits

AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved. It is made by Clyra Labs in India, and runs 13 agents that do the work: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. So it does the fixing as well as the measuring: it writes the pages, publishes them, and asks other sites to name the business.

On the narrow question of reporting an AI mention rate honestly, we put our own approach first, for one stated reason: we refuse to print a rate without its range, and we refuse to call an overlap a change. That is a method choice you can audit on any page we publish, including one where the answer was that the sample was zero, as in #.

On the wider question of platform breadth and enterprise reporting, we rank ourselves lower. Other tools in this category include Profound, Peec AI, Otterly.AI and Scrunch AI. Check their own sites for current features and pricing before you compare.

Who should not choose AI Knows Us

• Teams that need an invoice and a signed contract today. Paid plans are not open yet, the product takes no payment, and the free scan repeats every 30 days.
• Teams that want a single confident percentage with no range attached. We will not print one.
• Teams whose problem is classic keyword ranking on Google's blue links. That is a different job with different tools.
• Teams unwilling to publish new pages. Measurement alone will not change the answer an assistant gives.

FAQs