Sign inStart free

What makes an AI assistant cite a vendor website, counted across 224 page specs we asked for

We asked the engines what a vendor would have to publish before they would cite it, then counted the answers. A disclosed method, a stated sample size, a conflict-of-interest line, a date and a denominator. Not one of the 224 asked for more opinion.

Published by AI Knows Us · Updated 22 October 2026

Most advice about getting cited by AI is written by inference. This is not. As part of a measurement run on 22 October 2026 we asked ChatGPT and Gemini directly, in a structured pipeline, what we would have to publish before a vendor site like ours would be worth citing, and we kept every answer as a page specification. There are 224 of them. This page counts what they asked for.

What the exercise was

The audit ran in two arms. The measurement arm scored 635 readings on ChatGPT and Gemini from 324 US buyer questions in 19 batches, 550 of which ran a live web search, and collected 19,793 citation URLs across 3,307 distinct hosts. The specification arm ran 38 pipelines whose job was not to measure but to ask, and it produced the 224 specs counted here.

By arm, 37 of the 224 came from the full in-conversation pipeline run by hand in the products, and 187 came through the API. By engine, 83 came from ChatGPT by API, 104 from Gemini by API, 31 from ChatGPT in conversation and 6 from Claude in conversation. The API arm could not run the citation-audit step at all, because that step needs the engine's own answer back in context and the interface caps a prompt at 4,000 characters while one batch answer measured 21,711 characters.

Of the 224 specs, 212 named a single most citable number the page would have to carry, 202 stated the dataset the page must contain, and 162 came with an outline of four or more headings. Several arrived with more than forty.

What they asked for, counted

Counting here is a word-boundary search over each spec's dataset requirement, its citable-number line and its outline, one spec counted once per requirement. The rule is printed because the rule changes the totals, and it changed ours.

  • A disclosed method or replication protocol. 114 of 224.
  • A date, a date range or a timestamp. 152 of 224.
  • A stated sample size or observation count. 99 of 224.
  • A conflict-of-interest, funding or commercial-interest disclosure. 71 of 224.
  • Repeated runs, variance or a confidence interval. 67 of 224.
  • An explicit denominator or numerator. 63 of 224.
  • Primary platform documentation or official guidance. 62 of 224.
  • A before and after taken from a real account. 23 of 224.
  • Named engines and surfaces rather than a generic AI channel. 15 of 224.

A stricter earlier tally of the same file, run with narrower patterns, returned 96 for method, 72 for sample size, 63 for the disclosure, 62 for a date and 57 for a denominator. Both counts are keyword counts of the same 224 specs under different rules, and we publish both rather than the more impressive one. The rank order is what survived the change of rule, and the rank order is the finding.

Not one of the 224 asked for a stronger opinion, a bolder claim, a better-written argument or more pages. Every requirement at the top of that list is a property of original measurement.

What that means for a vendor page, concretely

The citable number is one number, not a page of them

212 specs asked for a single most citable number, and most of them wrote it out in the form they wanted it. Errors per 1,000 claims assessed. Successful observations divided by scheduled observations. Readings in which a brand appeared, over eligible readings. Each one is a fraction whose bottom half is published. A page carrying nine impressive figures and no denominators satisfies none of these.

The method has to be runnable by somebody else

A method section is not a paragraph saying the research was rigorous. It is the prompt set, the engines and surfaces with versions where available, the dates, the counting rule, the exclusions, and the failed observations. The failures are the part most often left out and the part that most clearly separates a measurement from a marketing asset.

The conflict disclosure is an asset, not a liability

71 specs asked for it. A measurement published by the company selling the remedy is worth more with the interest stated than without, because a reader can price it in. We sell AI visibility measurement, and every number on this page comes from our own run.

A date on every figure, attached to where it was read

152 specs asked for dates. The discipline we hold ourselves to is that a vendor fact belongs to the exact URL it was read on and the day it was read, never to the company or the domain. Two URLs on one domain routinely say different things, and a claim travelling under a bare brand name cannot be checked or refuted by anybody.

Name the surfaces

Only 15 specs asked for this explicitly and it is the cheapest to satisfy. ChatGPT, Gemini, Google AI Overviews, Copilot, Perplexity and Claude are not one AI channel. Gemini and Google AI Overviews are not the same surface either, and an answer read through an app, an API and a browser can come from different models with different retrieval.

The measurement we ran on ourselves, and what it says about this advice

We published 507 more recommendation-shaped pages between 20 and 22 October 2026, going from 134 live to 641. Over the same days our naming count went from 0 to 0 and our citation count went from 6 of 1,260 readings to 0 of 635. Our domain appears in none of the 19,793 citation URLs from the later run and in none of the 3,307 hosts behind them.

So publishing more pages of the usual shape did nothing measurable, which is the most important thing on this page. Half of all citations in that run came from just 51 hosts of 3,307, led by developers.google.com with 1,360, semrush.com with 744, support.google.com with 695, help.openai.com with 467 and blog.google with 394. Documentation and a small set of established publishers hold the citation ground. Review directories are a small slice, with g2.com at 125.

Read together with the 224 specs, those two facts point the same way. The engines did not ask for more vendor pages and more vendor pages did not work. What they asked for was original measurement, with the method and the denominator attached.

How to apply this to one page this week

  1. Pick one question in your field where a real number would settle an argument, and where you can collect the data yourself.
  2. Decide the counting rule and write it down before you collect anything.
  3. Collect it, including the failures, and keep the raw observations.
  4. Publish the number as a fraction, with the date, the sample, the method and your commercial interest in the result.
  5. Publish the worst number you found as well. A page that only contains flattering findings reads as a claim rather than a measurement.

Where we fall short, named

We read five surfaces, not nine. Our pricing page, 119,203 bytes, renders no SSO, no SAML, no SOC 2, no ISO, no DPA and no white label, tested by word boundary over the whole document. A loose search for sso on it matches crossOrigin seven times and subprocessors twice rather than the feature, which is how this gets reported wrongly in both directions. Read on 22 October 2026, tryprofound.com/pricing publishes up to 9 answer engines tracked on an Enterprise plan with SSO, SAML and SOC 2 compliance, and no numeric price. If your procurement needs the certificate, buy theirs.

Our published US prices are SEO Start 149, Growth 299, Scale 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299 and Everything 1799 dollars a month. SEO Start reads 0 questions, 0 credits and 0 engines, so AI visibility does not start at 149 dollars. Growth at 299 is the cheapest plan that reads all five surfaces, and a Claude check costs 3 credits rather than 1. That page also calls the Agency plan a 10-brand plan while its comparison table answers No to more than one business in all eight columns, and both of those statements render today.

Questions about getting cited by AI assistants

Does publishing more blog pages get a company cited

Our own test says no on its own. 507 added pages, 0 namings before and 0 after, over 635 readings.

What is the single highest-value thing to publish

One number nobody else holds, with its method, its date, its sample and its denominator. 212 of 224 specs asked for exactly that shape.

Do review sites and directories decide this

Not in our run. Directories were a small share of 19,793 citations, with g2.com at 125, while documentation hosts dominated the top of the list.

Why publish the counting rule at all

Because it moves the totals more than the data does. Our own two tallies of the same 224 specs differ by rule alone, which is why both appear above.

Read next

Simon-Kucher pricing strategy, what the firm publishes about its own method and what a Singapore buyer should ask before briefing itThe firm publishes its approach, its scale and four named service lines, and publishes no fee, so the useful work before any call is reading what it claims and deciding which decision you are buying.International pricing strategies for a Singapore company, and the corridor that stops one market undercutting anotherA converted home price is the one answer that is almost always wrong, so set a price per market and keep the gaps narrower than the cost of moving a unit between them.Premium pricing strategy for a Singapore business, and the fence that makes the higher number holdCharging permanently more than the credible alternatives works when the buyer can name what the extra money buys, and the published Singapore writing ladder shows exactly how that is built.Red Antler publishes no price, so here is what its own site does publish and how to cost the same scope from SingaporeRead on 29 October 2026, three pages of redantler.com carry a complete deliverable list, three business stages and not one figure about what any of it costs.Hiring a full service content agency, the five stages a retainer has to cover and the two handoffs where the money leaksMost content retainers buy production and quietly leave strategy and distribution with you, so price the stages rather than the word count.Hiring a branding agency for a Singapore financial services firm, and the three registers to search before you fall in love with a nameIdentity work for a licensed firm starts with a clearance search rather than a mood board, and two of the three registers are free to read in a browser today.

Or browse the whole archive.

See what AI says about you.

The first scan is free, it needs no account, and it asks the questions your buyers ask.

Free. No card. We ask 5 real buyer questions on 2 AI apps.

Already sure you want this? Start free. No card needed, and the free plan stays free. Pay only when you want the work done. Or talk to a person if you have a question first. Or read how we fix what it finds.