905 live-answer fetches in 30 days. That is how often ChatGPT fetched two groas pages during live retrieval, not how often it recommended us.
I pulled the logs because I am tired of watching one visibility number do four different jobs. Across those pages, 1,333 bot-matched fetches narrow to 1,312 verified fetches, then to 905 live-answer fetches. Each step answers a narrower question. None answers the question a business actually cares about: did this turn into a recommendation, a lead or revenue?
The distinction is not accounting trivia. It determines what an outsourced SEO and AI-search provider can honestly claim to have delivered. Here is what our first-party evidence establishes, where it stops, and the reporting split I would require before buying the work.
The same traffic produces three different totals
The headline total comes from two related Google Ads update pages measured over the same 30-day window. The main page is the one we published in full; the second uses the same logging setup. Both pass through the same sequence: User-Agent matching, IP verification, then live-answer classification.
| Page | Bot-matched fetches | Verified fetches | Verified ChatGPT live-answer fetches |
|---|---|---|---|
| Google Ads updates page | 850 | 833 | 513 |
| Second updates page | 483 | 479 | 392 |
| Combined | 1,333 | 1,312 | 905 |
On the main page, the first filter removes 17 requests. On the second, it removes 4. Those requests carry a matching bot label but do not pass IP verification. I do not count them as verified OpenAI traffic.
The pages also attract different mixes of activity. Live-answer fetches account for roughly 62% of verified requests on the first page and 82% on the second. Combining them hides that difference.
There is a mechanism behind the split: OpenAI uses separate crawlers for separate jobs. OAI-SearchBot handles search indexing; ChatGPT-User fetches pages in response to live requests. More background crawling and more live retrieval are not interchangeable signs of progress.
Read the combined row from left to right. Verification removes 21 requests that do not check out. Classification sets aside another 407 verified requests associated with indexing or training collection rather than live-answer support.
The largest total is not the most useful total. The useful number depends on the question being asked.
Verification establishes who fetched, not who saw you
A User-Agent string is a claim, not proof. Any script can put GPTBot or ChatGPT-User in a header. I used to treat a bot hit as a bot hit. I was wrong.
Our verification step checks the request’s source IP against OpenAI’s published IP ranges, rather than trusting the label. Requests that fail that check stay outside the verified total. That does not establish who sent them; it establishes that they do not meet our verification rule.
Then comes a separate classification question: what kind of request was it? A verified indexing crawl is still not a verified live-answer fetch. Lumping them together gives a precise-looking answer to an undefined question.
Even a verified live-answer fetch does not establish visibility in the final response. In an analysis of 15,000 ChatGPT search prompts, 548,534 pages were retrieved, but only 15% were cited. That study shows the gap between retrieval and citation. It does not tell me which of our fetched pages appeared in an answer, and I cannot apply its citation rate to our logs to manufacture an estimate.
Verification establishes access. Citation requires observing the answer. If a provider reports bot traffic without describing its verification and classification methods, I treat the total as unfiltered traffic, not demonstrated visibility.
What our live-answer total actually proves
The 905 verified live-answer fetches establish something narrow and useful: ChatGPT retrieved these pages during live requests in the measurement window. Our content was accessible through that retrieval path.
I care about that. It is first-party evidence of access, not a screenshot or a third-party visibility estimate. But the log does not contain the final answer, and it does not reveal the question that triggered the request.
Here is the vocabulary I use when reviewing reports:
| Event | What I mean by it |
|---|---|
| Fetched | ChatGPT retrieved the page. |
| Cited | The page appeared as a linked source in the answer. |
| Mentioned | The answer named the brand, with or without a link. |
| Recommended | The answer suggested the brand as an option. |
| Converted | A visit became a business outcome, such as qualified pipeline or revenue. |
These are distinct events, not different labels for the same achievement. A citation is not automatically a recommendation. A brand mention is not automatically a click. A click is not automatically a qualified lead.
Nor are fetches unique users or unique answers. An answer can trigger multiple fetches, retries or prefetches. Because both pages cover Google Ads updates, the same answer could retrieve both. Our logs count requests per URL; they do not de-duplicate conversations.
I can defend retrieval volume. I cannot turn it into a count of people reached. Anyone presenting the combined total as separate encounters with the brand has crossed beyond what these logs establish.
Citations and mentions need their own observations
Even if our log total represented citations, I would still want dates and repeat runs before calling it a trend.
Ahrefs tracked 11,225 sites cited by ChatGPT on the first day of a 20-day study. Its findings make a single screenshot look considerably less impressive:
| Finding | What it means for reporting |
|---|---|
| 59% were gone by the next day | A citation observed today may not persist tomorrow. |
| 4.0% lasted all 20 days | An initial appearance does not establish sustained presence. |
| About 75% of lost citations returned at some point | Disappearance does not necessarily mean permanent loss. |
Those are findings from that study, not persistence rates for groas. Their value here is methodological: a citation count without observation dates and repeat runs is a snapshot pretending to be a trend.
I also keep mentions separate. In the same research, 54% of ChatGPT citations did not name the brand in the answer text. Losing a citation while retaining the mention occurred about 13% of the time day to day.
A linked source and an answer that names the brand as an option are different outcomes. Both may matter. Neither deserves to disappear inside a blended visibility score that prevents me from seeing which one changed.
My reporting rule is simple: capture the answer, record what appeared, and keep citation, mention and recommendation labels separate.
Where the dataset stops
The published path list covers selected pages, not the entire site. It establishes what those pages received. It does not establish what every other page missed.
A page absent from that list is not automatically a content gap. Fetch leaders may attract requests because they are fresh, fast or interlinked. A different page may remain unfetched for reasons the list cannot diagnose. I use the logs to identify observed activity, not to invent an explanation for every absence.
The same limit applies to intent. The server log records the fetch, not the prompt. Without observing the questions, I cannot claim the requests came from buying queries, branded queries or a particular stage of the customer journey. Knowing the page’s subject does not let me reconstruct the conversation.
Engine coverage stops at the same boundary. These are ChatGPT-User requests verified against OpenAI ranges. They do not establish visibility in Gemini, Claude or Perplexity. Those systems have their own crawlers, retrieval logic and citation behavior. I do not let one engine’s evidence become an all-engine headline.
This is also why AI-visibility tools can report different results for the same brand. They may observe different surfaces and conditions:
- API responses versus browser-captured answers.
- Fixed prompt lists versus dynamic follow-ups.
- Different engines, geographic locations and refresh schedules.
An API response is not necessarily the same observation as the answer a person sees in a browser. Before arguing over which score is right, I want to know what each tool actually measured.
The practical limit is not embarrassing. It is useful. A measurement with a clear boundary tells me what to investigate next; an all-purpose score mostly tells me the reporting software works.
The reporting split I require before outsourcing
When I evaluate an ongoing SEO and AI-search provider, I want the evidence separated before I look at the headline. My minimum reporting structure is this:
| Reporting layer | Evidence I expect | What it must not be presented as |
|---|---|---|
| Verified access | Server requests, IP verification and crawler classification. | People reached or recommendations earned. |
| Observed answers | Captured citations and mentions, labeled separately. | Automatic endorsement or business value. |
| Recommendations | Recorded suggestions across repeat runs on a defined prompt panel. | Guaranteed placement or universal visibility. |
| Business outcomes | Attributed calls, qualified pipeline or revenue. | A synonym for traffic or answer appearances. |
That last separation matters. An analysis covering 100 sites and 145 million sessions found ChatGPT traffic equal to 0.62% of Google organic volume, with all LLMs at 0.73%. Google drove 96.5% of conversions in that dataset.
Those figures describe that sample, not every business. They do not make AI-search work pointless. They do make it hard to defend a report that celebrates visibility while leaving acquisition and conversion outcomes out of view.
I want to see where activity becomes business value, not have the report assume the connection. A provider can show more fetches, citations or mentions while pipeline stays flat. Keeping the layers separate makes that visible.
Record the method, then name who ships the fix
The reporting table is only useful if someone can repeat the observation. Alongside it, I require a written method:
- The exact prompt panel.
- The engine used for each run.
- The observation date and repeat schedule.
- The captured answer and the labels applied to it.
- The changes made in response to the findings.
I run ChatGPT, Gemini and Perplexity separately rather than combining them into an engine-agnostic result. For a working version of that routine, I use this method for building a prompt panel and tracking visibility. The point is a repeatable log, not a folder of flattering screenshots.
Then I ask the question reporting decks tend to postpone: who changes the site?
If the work stops at identifying a problem, I am buying diagnosis without implementation. I want publishing, technical fixes and content edits assigned to someone who can deliver them. A report without that is a weather forecast.
Buy the edits, not the chart. The method should tell me what happened; the operating model should tell me who does something about it.
Buy implementation, not a larger visibility number
What I want from groas is the same evidentiary split, attached to done-for-you execution. groas runs paid ads, SEO and AI-search work under a model where I set goals and budget and a senior owner answers for the result. The work comes with the reasoning behind it, in plain language, rather than another dashboard I have to interpret.
That means verified fetches belong in their own column. Observed citations, mentions and recommendations need their own records. Qualified pipeline, closed-won revenue, ROAS and cost per acquisition remain the business measures, not decorative additions to a visibility report.
If answer observations reveal a problem, someone needs to investigate and make the appropriate change. The fetch log alone cannot tell that person whether a page was cited or discarded. Keeping those evidence sources separate is how I avoid commissioning a rewrite for a problem the logs never established.
The advantage I want is accountable execution, not a more flattering definition of visibility. No monthly PDF that calls every bot hit a win. No recommendation implied by a retrieval count.
No provider can promise ChatGPT’s recommendation, groas included. No one controls the selection step, and no ChatGPT log proves anything about another engine. So I judge the hire by the reporting and the shipping: separate the evidence, record the method, then publish the fix when the observations support it.
The 905 fetches tell me our pages were retrieved. The decision they support is not to declare victory. It is to buy the team that can measure the next step honestly and do the work to improve it.

