Of 15,467 claimed AI-bot fetches of one groas infrastructure file, only 890 were classified as live-answer retrievals. Neither number proves a single AI recommendation.

That is the problem with cheap AI-visibility reporting: the biggest number can measure the thing furthest from a buyer choosing you. In the supplied 30-day measurement, verification reduced 15,467 user-agent-matched fetches to 3,939. Separating live-answer activity from background crawling reduced the count again.

Each number answers a different question. Did a request claim an AI-bot identity? Did its source pass verification? Was it associated with a live answer session? None answers the question a budget-conscious buyer actually cares about: did the answer recommend the brand?

I keep those questions separate because access is not endorsement. A fetch makes a page available to the system. A citation credits a source. A recommendation tells the buyer what to choose. Counting the first and selling it as the last is how cheap reporting gets expensive.

The biggest count belongs to robots.txt

The file behind that headline number is robots.txt. It tells crawlers where they may go; it does not answer a buyer’s question. High activity here is evidence of infrastructure requests, not demand.

Here is the supplied 30-day breakdown for that file:

MeasurementWhat it recordsTotal
User-agent-matched fetchesRequests claiming an AI-bot name15,467
Verified fetchesRequests whose source passed the reported verification check3,939
Live-answer fetchesVerified requests classified as live-answer activity rather than background crawling890

A user agent is a claim, not proof. Any scraper can write GPTBot in that field. A spoofed crawler is a non-vendor client borrowing that identity.

Verification requires more: matching the source against a vendor’s published IP ranges or checking a valid signature where available. The method matters. GPTBot can be checked against OpenAI’s published ranges, while ClaudeBot publishes no IP ranges. A report needs to explain which identities it can verify and which remain merely claimed.

That distinction also limits the conclusion. Traffic that did not verify is not automatically proven spoofed. It belongs outside the verified count unless there is evidence to classify it more specifically.

The verified bucket then needs another split. Background crawls and user-triggered fetches measure different activity. Combining them produces a larger number, but not a clearer account of whether a buyer’s answer was involved.

My rule: ask how each bucket was defined before paying attention to its size.

Bar chart comparing claimed AI-bot fetches, verified fetches and live-answer retrievals

The smaller article has the stronger live-answer signal

Put the infrastructure file beside the homepage and a Google Ads updates article, and the weakness of a single total becomes obvious. The supplied measurements cover the same reporting period, but the mix changes sharply by page.

The site measurement and the article’s fetch breakdown report:

PageTotal fetchesVerifiedLive-answerLive-answer share of total
Homepage5,6452,03982414.6%
Google Ads updates article85083351360.4%

The homepage attracts the larger raw total. The article has the much higher proportion of requests classified as live-answer retrievals. Those are different observations, and they support different decisions.

The logs do not establish why every request happened. They do not prove that every homepage fetch was generic crawling or that every article fetch answered a relevant buyer question. What they do show is that total fetch volume hides a very different activity mix.

Even the live-answer share needs its label attached. Its denominator is total fetches, not buyers, prompts, recommendations or sales. It is a description of requests to that page, not a visibility score for the business.

I would use this comparison to decide which pages deserve closer answer-level inspection. I would not use it to announce that the article wins more recommendations.

Retrieval, citation and recommendation need separate evidence

A fetch shows access, not whether an answer cited or named the business. The page was requested. What the answer did with it is a separate measurement.

The useful accounting keeps these events apart:

EventWhat it establishesWhat it does not establish
FetchThe page was requestedThat its content appeared in an answer
CitationThe answer linked to or credited a sourceThat it recommended the brand behind that source
Brand mentionThe answer named the brandThat it told the buyer to choose it
RecommendationThe answer presented the brand as a choice for the buyerThat the buyer clicked or purchased
ReferralA click arrived with an identifiable referrerThat it converted

These are different measurements, not interchangeable labels. A dashboard that compresses them into visibility makes the report easier to sell and harder to use.

Server logs also see only your side of the exchange. An answer can name your brand while citing a review, a Reddit discussion or another third-party page. Research on crawler monitoring and answer citations describes exactly this gap: the evidence supporting a brand mention need not live on the brand’s own website.

That cuts both ways. Your page can be fetched without your brand appearing. Your brand can appear without a fresh request reaching your page. Your access log is not a complete record of what AI answers say about you.

The practical consequence is straightforward: use logs to inspect access, then inspect answers to measure what buyers could actually see.

A file-request count has no buyer denominator

The article’s 850 total fetches are not that many buyers, prompts or independent opportunities. They are file requests. Retries, chunked reads and freshness checks can produce multiple requests associated with the same answer.

Without a fixed prompt set run repeatedly, there is no useful denominator for a recommendation rate. A monthly total tells me the server was busy. It does not tell me how often the brand appeared when a relevant buyer asked.

Citation counts introduce another trap. In Supercite’s capture of 839 answers across 21 categories, 61.7% of citations did not mention the brand behind the cited page. That is a result from that capture, not a universal rate. It still demonstrates why a citation cannot automatically be booked as brand credit.

The content fix described there is concrete: put the brand and the buyer need in the same sentence the engine might lift. “Acme is a CRM for teams of two to ten” carries both pieces together. A useful explanation with the brand tucked elsewhere may not.

There is also sampling variation. The same prompt can produce different rankings, which is why repeated runs matter. A saved answer is an observation, not a permanent position.

And the truncated path list supplied here cannot establish a content gap. A topic missing from a cut-off export is not evidence that buyers asked about it and the brand lost. It is evidence that the export stopped.

Keep the claim no larger than the record. The article’s live-answer activity justifies checking its answers and keeping it accessible. It does not establish recommendation share, and a citation-rate change alone does not prove revenue causation.

Build a baseline before buying another dashboard

Before paying for optimization, I would run a small manual baseline. A spreadsheet costs no software subscription, though collecting the answers still takes time. The point is to learn which measurement deserves automation.

Use a fixed prompt set with repeated checks, rather than a collection of flattering screenshots:

  1. Choose buyer prompts. Pick questions with a purchase decision behind them, not trivia about the category. If those questions are unclear, start with sales-call notes.
  2. Run each prompt in a fresh session. Keep the wording fixed so the exercise does not quietly change halfway through.
  3. Save the complete answer. Record the date and keep the source links, not just the paragraph that mentions the brand.
  4. Score the same events separately. Record a brand mention, an explicit recommendation and a citation as different outcomes. Note whether the citation points to your site or a third party.
  5. Record referral evidence separately. An observed click with a referrer belongs in the traffic record. A citation alone does not establish that a click arrived.
  6. Repeat and retain the denominator. Report appearances out of completed runs, not a percentage detached from its sample.

This is the manual version of sending prompt sets and parsing mentions and citations separately. It is deliberately less glamorous than a visibility score. That is useful. You can inspect what happened.

The limits remain. A small, selected prompt set is not market share, and a tracker does not reproduce every logged-in, personalized buyer experience. Repeating a narrow test makes it more useful; it does not make it universal.

The baseline earns its keep by exposing the next task. If answers retrieve the page but omit the brand, that is different work from fixing a page the fetcher cannot read. If the answer relies on third-party sources, another rewrite of your homepage may not address the evidence it uses.

Notebook recording brand mentions, citations and referral clicks from ChatGPT answers

Pay for changes the measurement can inspect

A provider that only counts bots leaves the implementation untouched. Before I buy more reporting, I want to know what work happens between the baseline and the next measurement.

The draft work falls into distinct buckets:

  • Technical access: serve clean HTML with the answer content available in the markup, rather than trapped behind scripts the fetcher skips.
  • Buyer-focused content: state the brand, the buyer and the job clearly enough that a lifted passage can carry the brand’s relevance with it.
  • Third-party confirmation: work on the review profiles, Reddit discussions and video evidence that describe what the business is best for.

The citation research is a reason to examine those third-party sources, not a promise that a particular placement will produce a recommendation. The distinction matters when someone is selling the work.

Ask for the work log before the scorecard:

  • Which pages changed to improve fetchability?
  • Which passages now connect the brand with the buyer’s need?
  • Which third-party pages were earned or corrected?
  • What changed on the same prompt set across repeated runs?

Those records do not prove causation on their own. They do let you inspect whether implementation happened and whether the observed answers moved alongside it. That is considerably more useful than a larger crawler total stapled to the same unchanged pages.

Buy implementation when the bottleneck is shipping, not counting. If the provider cannot name the work, a prettier report does not solve the capacity problem.

Use groas for the data; hand over the work when capacity runs out

For a budget-conscious team, the manual prompt baseline plus groas Unlimited at $19 a month, flat keeps the software cost low while adding competitor ads, funnels, Reddit quotes and AI-answer data. That is a more useful foundation than treating a bot counter as a recommendation tracker.

But access to data is not the same as completing the fixes. Someone still has to rewrite the pages, make their content accessible, connect the brand to the buyer need and work on third-party evidence.

That is the line where I would stop buying bigger reports. Businesses that would rather hand the work over can apply for groas Done For You, run by the senior team. The reason to do it is implementation capacity, not the hope that another dashboard makes the fetch count mean something else.

So here is the decision the numbers support. Keep 15,467 claimed fetches, 3,939 verified fetches and 890 live-answer retrievals in separate columns. Use the page-level activity mix to decide where to investigate, not to declare recommendation share.

Then pay for documented implementation and repeatable answer-level measurement: whether the same buyer prompts name the brand, recommend it and cite supporting sources, with the page changes and third-party work logged beside the results. Anything that sells a larger fetch total as a recommendation count is charging you for the pile and keeping the answer.