RESEARCH · MEASUREMENT

How AI citations are measured here

There is no equivalent of Search Console for AI assistants, so citation share is measured with a fixed prompt panel: a written set of buyer questions, run on a schedule across each assistant, with the answers recorded verbatim along with whether the brand was named and whether a link was given. The output is the prompts and the answers, not a proprietary score.

Last updated . Written and verified by Bhushan Raj Shakya, founder of RankedTag.

KEY FACTS

The problem
No assistant publishes a per-site impressions report. Server logs show crawler visits, not citations
The method
A fixed prompt panel, run on a schedule, answers recorded verbatim
Panel size
Typically 20 to 40 prompts, written once and frozen
What is recorded
Named or not, linked or not, position in the answer, competitors named alongside
Reported as
The prompts and the answers. No composite score
Known limits
Answers are non-deterministic, personalised and change without notice

Why this has to be measured by hand

Google Search Console reports impressions, clicks, position and query for Google Search Google Search Console Help, performance report. Nothing equivalent exists for ChatGPT, Claude or Perplexity: there is no per-site report of how often a brand was named in an answer, because the assistant operator is not running a search results page on your behalf.

Two proxies are often substituted and both are weak. Referral traffic from an assistant undercounts badly, because a buyer who reads a recommendation in an answer frequently searches the brand name afterwards and arrives as organic or direct. Crawler hits in server logs show that an agent fetched a page, which is a prerequisite for citation and not evidence of it. The operators document those agents precisely because fetching and citing are different events OpenAI, bots and crawler documentation.

That leaves direct observation, which is what a prompt panel is: ask the questions a buyer would ask, record what comes back, and do it the same way every time.

How the panel is built

  1. Write the prompts from buyer language, then freeze them

    Prompts come from sales calls, support tickets and the queries that already convert, phrased the way a person asks an assistant rather than the way they type into a search box. Once written, they do not change, because a panel that changes cannot show a trend.

  2. Cover the three question types that matter

    Recommendation prompts ("best X for Y"), comparison prompts ("X vs Y for a company like mine"), and verification prompts ("is X any good for Y"). Each behaves differently and a panel of only one type reads as a flat line.

  3. Run across assistants on a fixed schedule

    The same panel, the same wording, run across each assistant in scope on the same cadence, with no personalisation carried over between runs where that can be controlled.

  4. Record the answer verbatim

    The full answer text, whether the brand was named, whether a link was included, where in the answer it appeared, and which competitors were named alongside. Screenshots for anything that will be quoted publicly.

  5. Report the raw material

    The client receives the prompts and the answers. A composite score hides exactly the information that makes the number actionable, which is which answers moved and what they now say.

What actually moves the number

The published research is narrow but real. The peer-reviewed GEO study, which tested content changes against generative engines, found quotation-heavy content gained roughly 41% in visibility, and that adding cited sources produced the largest gain for a page at position five, around 115% Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024. Those findings point at the content itself rather than at any technical trick.

In practice the levers are: being crawlable by the agents in question Anthropic Support, does Anthropic crawl data from the web, making claims that can be quoted because they are specific and sourced, describing the entity consistently enough that it can be resolved, and covering the comparison and alternatives questions buyers actually ask, since those are the prompts that produce recommendation answers.

What this method cannot tell you

  • It is not a census. A panel of 30 prompts is a sample of a question space that is effectively unbounded.
  • Assistant answers are non-deterministic. The same prompt can return different answers minutes apart, which is why a panel is run on a schedule and read as a trend rather than a reading.
  • Personalisation and geography affect answers, and neither can be fully controlled from outside.
  • It measures citation, not revenue. Attribution from assistant answers to signups remains the hardest part of the chain, which is why the first-party sample published here is interesting but small RankedTag, 17 production signup alerts from the Sendr engagement.
  • It cannot prove causation. A panel shows what changed, in parallel with work that was done; it does not isolate variables the way a controlled experiment would.

Frequently asked questions

Can I see AI assistant traffic in Google Analytics?

Partially and unreliably. Some assistant referrals appear as referral traffic, but buyers frequently search the brand afterwards, so the visit arrives as organic or direct and the assistant gets no credit.

Do server logs prove an assistant cited my site?

No. Logs show that a documented agent fetched a page. Fetching is a prerequisite for citation, not evidence of it.

How many prompts should a panel have?

Typically 20 to 40. Enough to cover recommendation, comparison and verification questions, few enough that every answer can actually be read rather than summarised.

How often should the panel be run?

On a fixed cadence, commonly fortnightly or monthly. Consistency matters more than frequency, because the panel only has meaning as a trend.

Why not report a single visibility score?

Because the score hides the useful part. Knowing that three recommendation answers now name you, and what they say, is actionable. Knowing that a composite rose from 24 to 29 is not.

Official resources

FOUR SLOTS A MONTH

Measurement you can read.
Not a score you have to trust.

The 52-check founder review is run personally and comes back within 48 hours, with a real opinion even when the answer is no.