---
title: "905 Live-Answer Fetches, Not 905 Recommendations: Reading an AI Search Report"
description: "Two groas articles logged 905 verified live-answer fetches in 30 days. See what those numbers prove—and what an AI search report must show next."
url: "https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/64a17442-4342-43aa-ad62-e5a9c4128682.png"
published: "2026-11-07T06:18:38.803Z"
modified: "2026-11-07T06:18:38.858Z"
---

November 7, 2026 · 9 min read

# 905 Live-Answer Fetches, Not 905 Recommendations: Reading an AI Search Report

[](https://groas.com/author/alexander-perelman)

[Alexander Perelman](https://groas.com/author/alexander-perelman)Personally Tested

Head Of Product @ groas

[LinkedIn](https://www.linkedin.com/in/alexander-433793253/)

![A miniature wooden ladder with tiny figures swarming its bottom rung carrying papers, while the upper rungs and a dusty cash register on top stay empty.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/64a17442-4342-43aa-ad62-e5a9c4128682.png)

In this article

1. [Two articles account for 905 live-answer fetches](#two-articles-account-for-905-live-answer-fetches)
2. [The busy RSS feed shows why verification comes first](#the-busy-rss-feed-shows-why-verification-comes-first)
3. [Four evidence stages—and what each one proves](#four-evidence-stagesand-what-each-one-proves)
4. [What this dataset leaves unknown](#what-this-dataset-leaves-unknown)
5. [Ask for three reports, not one blended score](#ask-for-three-reports-not-one-blended-score)
6. [Use consumption to choose the next edit—not the next batch](#use-consumption-to-choose-the-next-editnot-the-next-batch)
7. [The same reporting standard applies to groas](#the-same-reporting-standard-applies-to-groas)

[905 live-answer fetches in 30 days](https://groas.com/post/what-30-days-of-ai-bot-traffic-on-our-own-site-showed-two). That is how often verified systems fetched two Google Ads update articles on groas’s site while building answers—not how often they recommended us.

I count that as **consumption, not wins**. The pages were retrieved in response to user questions. The log does not tell me whether an answer cited us, recommended us, or sent us a qualified inquiry.

That distinction is the whole reporting problem. A verified fetch establishes access. A live-answer fetch establishes consumption. Neither establishes preference or revenue. Yet a report can slide between those claims so smoothly that a busy client never notices the missing evidence.

I want to follow the numbers one rung at a time: access, consumption, answer inspection, then money. Here is what groas’s measurements support—and where they stop.

## Two articles account for 905 live-answer fetches

I started with the two paths that looked like winners in our edge log. Both are Google Ads update articles. Together, they recorded the following:

| Path type                     | Total hits | Verified hits | Live-answer hits |
| ----------------------------- | ---------: | ------------: | ---------------: |
| Google Ads update article 1   |          — |             — |              513 |
| Google Ads update article 2   |          — |             — |              392 |
| Both update articles combined |      1,333 |         1,312 |              905 |

![Total bot hits and verified live-answer fetches for two Google Ads update articles](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/7ea8a8ff-aa08-4485-9f32-e3d590704368.png)

_Verified_ means the request IP matched the provider’s published ranges, not merely that its User-Agent claimed to be a known bot. The dashes mean the table does not provide those individual totals; they do not mean zero.

Verified requests account for **98.4% of recorded hits** on these paths, calculated from the combined totals above. Almost everything counted there was genuine provider traffic.

The verification method matters. I matched the true client IP behind our CDN against the [published IP ranges for OpenAI’s bots](https://en.natielimelech.com/blog/openai-bots-guide). Perplexity requires the same kind of check against its [separate crawler IP files](https://mbsearch.co/blog/perplexity-crawlers/).

That gives me confidence in who made the requests. It does not turn requests into people, answers, or recommendations. One user question can trigger several fetches. A fetch can produce no citation. A citation can name a business without recommending it.

**Read the table as verified retrieval activity, not a count of prospective customers.**

## The busy RSS feed shows why verification comes first

The same log contained a path that looked busy until I filtered it: our RSS feed.

| Path type                 | Total hits | Verified hits | Live-answer hits |
| ------------------------- | ---------: | ------------: | ---------------: |
| RSS / infrastructure path |        545 |             3 |                0 |

Report only the first column and this looks like activity worth celebrating. Read across, and there are barely any verified requests and no recorded live-answer fetches. That is why I argue that [answer-time fetches beat raw bot counts](https://groas.com/post/we-logged-30-days-of-ai-bot-traffic-on-o).

The mechanism is not complicated: [anyone can fake a User-Agent](https://en.natielimelech.com/blog/openai-bots-guide). A familiar crawler name is a claim, not identification.

A separate origin-log study illustrates the problem. Across [three sites and 16,548 requests claiming known crawler names, only 49.5% verified as genuine; 39.3% were spoofed](https://easybyte.es/lab/studies/s2/paper.pdf). Among requests claiming to be user-initiated fetchers such as ChatGPT-User and Perplexity-User, [67.8% were spoofed](https://easybyte.es/lab/studies/s2/paper.pdf).

Those are findings from that study, not a spoofing rate to apply to every website. Our update pages and RSS feed already show how differently paths can behave within the same log.

I used to treat a bot label as a visit. I was wrong. **Without verification, I am counting strangers in costumes.**

## Four evidence stages—and what each one proves

I use an _evidence ladder_ because each stage answers a different question. Moving up requires more evidence, not a more enthusiastic caption on the same chart.

| Stage                   | Question answered                                         | Evidence needed                          | What remains unproven                     |
| ----------------------- | --------------------------------------------------------- | ---------------------------------------- | ----------------------------------------- |
| Verified access         | Can the system retrieve the page?                         | A verified fetch                         | Use in a live answer                      |
| Live-answer consumption | Did a verified system retrieve it while answering a user? | A verified live-answer fetch             | Citation or recommendation                |
| Answer inspection       | Was the business cited, named, or recommended?            | Recorded answers from a fixed prompt set | Qualified demand or revenue               |
| Commercial measurement  | Did a qualified inquiry or sale follow?                   | Business outcomes tied to the activity   | Anything the attribution cannot establish |

### Access and consumption are different jobs

OpenAI separates its bots by purpose: [GPTBot collects training data, OAI-SearchBot crawls for search, and ChatGPT-User fetches after a user asks](https://en.natielimelech.com/blog/openai-bots-guide). Perplexity makes a similar distinction between [PerplexityBot for indexing and Perplexity-User for user-initiated retrieval](https://mbsearch.co/blog/perplexity-crawlers/).

That distinction is why I do not blend every verified request into an “AI visibility” total. An index crawl and an answer-time fetch tell me different things.

Our combined live-answer count clears the consumption stage. **A verified system retrieved the pages while building answers.** Useful evidence. Still not a recommendation.

### Answer inspection requires looking at answers

Server logs cannot tell me whether groas appeared as a source, a passing mention, or the recommended option. For that, I need to inspect answers to a fixed prompt set, repeat the exercise on a schedule, and record citations, mentions, recommendations, and share of voice.

This is part of what [ongoing GEO services describe as baseline audits and scheduled prompt re-runs](https://www.useomnia.com/blog/generative-engine-optimization-services), alongside content rewrites, schema and entity work, and digital PR.

The important word is _observed_. A recommendation in a recorded answer is evidence about that answer. It is not permission to relabel every fetch as a recommendation—or assume every user sees the same result.

**Keep the answer record beside the visibility claim.**

### Commercial measurement needs its own evidence

Even an observed citation does not establish a visit, much less a sale.

In Pew’s study of [68,879 searches, users clicked a traditional result 8% of the time when an AI summary appeared, versus 15% without one; about 1% clicked a link inside the summary](https://www.erste.company/zero-click-search/).

Those figures describe behavior in that study. They do not supply a conversion rate for groas, ChatGPT, Perplexity, or any individual page. I include them because they expose the leap from “appeared in an answer” to “generated demand.” The click itself cannot be assumed.

For commercial reporting, I want qualified inquiries, pipeline, and closed-won revenue—not a consumption number wearing a revenue label. **A fetch is not a citation, and a citation is not pipeline.**

## What this dataset leaves unknown

The source is a truncated list of paths that registered hits. That sounds like a footnote until someone turns an omitted topic into a content strategy.

**A topic missing from the list is not a verified gap.** It might have no demand behind it, sit outside the exported paths, or simply have gone unfetched during this measurement window. The list cannot distinguish those explanations.

Before accepting an “invisible topic” diagnosis, I would ask for:

- The full path export, not just the busiest pages.
- The query or prompt set expected to surface the topic.
- The answers observed when those prompts were run.
- The measurement window attached to the finding.

Fetch count also says nothing by itself about reach, sentiment, or money. A page used as background reading and a page used to recommend a provider can both generate a fetch. The log records retrieval, not the business value of the role the page played.

Finally, this is **one site over one 30-day window**: groas’s own. It is a worked example of how to read the metric, not an industry benchmark or a target for another business to copy.

I would use these numbers to choose what to inspect next. I would not use them to declare market-wide success.

## Ask for three reports, not one blended score

If I am paying for ongoing SEO and AI search, I want **verified crawling, observed answers, and business outcomes kept separate**. The reports can sit together. Their claims cannot be blended.

| Report            | What I expect to see                                                                        | What I will not accept as a substitute |
| ----------------- | ------------------------------------------------------------------------------------------- | -------------------------------------- |
| Verified crawling | IP-verified fetches by page, split into index crawls and live-answer fetches                | Raw User-Agent counts                  |
| Observed answers  | Fixed prompts, recorded answers, citations, mentions, recommendations, and share of voice   | An inference drawn from server logs    |
| Business outcomes | Qualified inquiries, pipeline, and closed-won revenue tied to the activity where measurable | A visibility score relabeled as ROI    |

I also want the limits written beside the results. If attribution is incomplete, say so. If a prompt set covers only part of the buying journey, show the prompts. If the export excludes paths, identify the coverage.

This is not a request for a more decorative dashboard. It is a request to know which question each metric answers.

**A provider should be able to explain the distance between its evidence and its claim.** If that explanation disappears into “AI-powered visibility,” I still do not have an answer.

## Use consumption to choose the next edit—not the next batch

Our most-consumed pages were Google Ads update articles: timely, dated material an assistant could retrieve when someone asks what changed. That makes sense as a consumption pattern. It does not prove buyer intent.

A prospect choosing who should manage search asks different questions: who owns the outcome, how quickly changes ship, and what it costs. If a heavily fetched page answers none of those, **I would not clone it just because the fetch count looks good**.

I would check the gap first:

1. Review recent sales calls or demo transcripts and list the recurring objections.
2. Open each high-consumption page and check whether it addresses those objections.
3. Inspect the relevant answers to see how the page is being used.
4. Rewrite the page to connect the information to the buying decision, or build a handoff to a page that does.

A dated update can help an assistant explain what changed. It may do little to help a buyer decide who should run search. Those are different content jobs. I want the handoff between them before commissioning more of the first.

That is the practical point behind [what the smaller, answer-time number should force a provider to report](https://groas.com/post/we-logged-30-days-of-ai-bots-on-our-own-3): a fetch identifies where retrieval is happening. It does not finish the commercial work.

![Ink cartoon of a marketer celebrating bot traffic while overlooking the citation that drives pipeline](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/5afdfdcd-1795-4da8-8a49-81f78547086d.png)

I prioritize documented breaks over rising volume:

1. **Verified access failures:** a known assistant cannot retrieve a page it should be able to access.
2. **Consumption without observed recommendations:** pages get live-answer fetches, but the fixed prompt set shows no citations or recommendations.
3. **Citations without qualified inquiries:** answers include the business, but commercial results remain missing.
4. **Bot volume:** useful context after verification, not the objective.

Each problem needs different work. More content does not repair an access failure. More fetches do not prove a recommendation. **Fix the broken rung before celebrating activity below it.**

## The same reporting standard applies to groas

If I sign this report, the method belongs on the first page:

- **Source:** groas’s own edge log.
- **Verification:** true client IP matched against provider-published ranges.
- **Coverage:** an export of paths with recorded hits, not proof that omitted topics are absent.
- **Window:** the measured period on this site, not a benchmark for everyone else.

That is the standard I expect from groas’s commitment to [explain the reasoning behind its actions](https://groas.com/our-philosophy). Our own report does not get a pass because the numbers look encouraging.

The finding is narrow: 905 live-answer fetches across two update articles establish access and consumption. Reach, preference, and revenue remain unproven. The next measurement is answer inspection: run fixed prompts on ChatGPT, Gemini, and Perplexity, record what appears, then trace qualified inquiries where possible.

The next content decision is equally specific. Connect those updates to the buyer’s question—who should run search when answers change this fast—or provide a clear handoff to a page that does. Do not commission another batch simply to raise the fetch count.

That is where I separate reporting from execution. groas exists to execute continuously and answer for pipeline, closed-won revenue, and CPA, not to make a monthly bot chart taller.

**Pay for movement down the ladder, not a higher hit total.** Ask the current provider for verified fetches, observed answers, and business outcomes in separate tables, with the gaps stated plainly. If they cannot distinguish what they measured from what they are selling, replace them.

## Frequently Asked Questions

### Does 905 live-answer fetches mean an AI assistant recommended groas 905 times?

No. The 905 figure counts verified systems retrieving two Google Ads update articles while building answers. It establishes access and consumption, but the log does not show whether an answer cited groas, recommended it, or produced any inquiry or revenue.

### How were the AI bot requests in the groas report verified as real?

The true client IP behind the CDN was matched against IP ranges published by the providers, such as OpenAI's published ranges for its bots. A User-Agent string alone is not proof, since anyone can fake a familiar crawler name. On the two update-article paths, verified requests made up 98.4% of recorded hits.

### Why should raw AI bot hit counts not be trusted?

A crawler name in a User-Agent is a claim, not identification. In the groas log, an RSS path recorded 545 total hits but only 3 verified requests and 0 live-answer fetches. A separate study across three sites found only 49.5% of 16,548 requests claiming known crawler names verified as genuine, with 39.3% spoofed.

### What evidence proves an AI assistant is actually recommending your business?

You need observed answers, not server logs: run a fixed prompt set on a schedule and record citations, mentions, recommendations, and share of voice. Even that is not enough for revenue reporting, which requires qualified inquiries, pipeline, and closed-won revenue tied to the activity.

### If a topic is missing from an AI visibility report, is that a content gap?

Not necessarily. The groas export was a truncated list of paths that registered hits, so a missing topic could have no demand, sit outside the exported paths, or simply not have been fetched during the 30-day window. Before accepting a gap diagnosis, ask for the full path export, the prompt set, the observed answers, and the measurement window.

### Do AI search summaries reduce clicks to websites?

In Pew's study of 68,879 searches, users clicked a traditional result 8% of the time when an AI summary appeared, versus 15% without one, and about 1% clicked a link inside the summary. These figures describe that study's behavior; they are not a conversion rate for any individual business or page.

### Should I create more content like my most-fetched pages to get more AI traffic?

Not automatically. The most-fetched groas pages were dated Google Ads update articles, which assistants retrieve for what-changed questions but which may not address buyer objections like who owns the outcome, speed, and cost. Check the gap against sales objections first, and fix broken rungs like access failures or citations without inquiries before chasing volume.

### What should I ask an AI search or GEO provider to report?

Ask for three separate reports: IP-verified fetches by page split into index crawls and live-answer fetches; recorded answers to a fixed prompt set with citations, mentions, recommendations, and share of voice; and qualified inquiries, pipeline, and closed-won revenue where measurable. Their claims cannot be blended, and a visibility score relabeled as ROI is not a substitute for business outcomes.

## Run Your Marketing On groas

Plug ChatGPT, Claude or any AI agent into groas and let it do the work of a whole marketing team: $19 a month flat with every feature included, and autopilot whenever you’re ready.

[Start For $19](https://groas.com/pricing)

[Or have groas run it for you](https://groas.com/done-for-you)

## Structured data

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","description":"groas is the marketing infrastructure for your own AI: it connects your AI to all your marketing and gives it proprietary marketing data collected 24/7. Plug Claude, ChatGPT or any AI agent into groas and run all your marketing from it, for a tiny fraction of the usual cost, with the groas app there whenever you want to see more. It is the layer ChatGPT, Claude and any AI assistant run a business’s online marketing on. Through groas an assistant can do anything you could do natively in Google Ads, Bing Ads, Meta Ads and ChatGPT ads, on your website and in AI search, and, where the owner switches it on, can run connected campaigns by itself, so AI agents do the work a whole marketing team used to. groas also gives those agents large-scale proprietary marketing data it collects around the clock, on any schedule the owner chooses: every ad competitors run on Google and Meta, their landing pages and full funnels with the offers and prices in them, Reddit customer research, AI search answers from ChatGPT, Google AI Overviews, Gemini, Claude and Perplexity, and SEO data from Semrush and Ahrefs, backlinks included ($19 a month, flat, with everything included, for unlimited teammates on one shared workspace and unlimited ad accounts and websites, delivered as an MCP server). Separately, its done-for-you service has senior people and proprietary quantitative systems run paid ads, SEO and AI search for businesses, and white label for agencies, from $5,000 a month.","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#webpage","url":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside","name":"905 Live-Answer Fetches, Not 905 Recommendations: Reading an AI Search Report","description":"Two groas articles logged 905 verified live-answer fetches in 30 days. See what those numbers prove—and what an AI search report must show next.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/64a17442-4342-43aa-ad62-e5a9c4128682.png"},"breadcrumb":{"@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#breadcrumb"},"dateModified":"2026-11-07T06:18:38.858Z","mainEntity":{"@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"905 Live-Answer Fetches, Not 905 Recommendations: Reading an AI Search Report","item":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#article","url":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside","mainEntityOfPage":{"@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"905 Live-Answer Fetches, Not 905 Recommendations: Reading an AI Search Report","description":"Two groas articles logged 905 verified live-answer fetches in 30 days. See what those numbers prove—and what an AI search report must show next.","datePublished":"2026-11-07T06:18:38.803Z","dateModified":"2026-11-07T06:18:38.858Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/64a17442-4342-43aa-ad62-e5a9c4128682.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"verified, live-answer, 905, live-answer fetches, fetches, answers, search, report","wordCount":1969,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/alexander-perelman#person","name":"Alexander Perelman","jobTitle":"Head Of Product @ groas","description":"My background includes Goldman Sachs and Stanford Computer Science.","image":"https://groas.com/media/blog/4f17cac3a81acc224cc1b2cfab8f94bbc295a3089d13e8421125771ef8d4064f.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"},"sameAs":["https://www.linkedin.com/in/alexander-433793253/"]},{"@type":"FAQPage","@id":"https://groas.com/post/905-live-answer-fetches-are-not-905-recommendations-inside#faq","mainEntity":[{"@type":"Question","name":"Does 905 live-answer fetches mean an AI assistant recommended groas 905 times?","acceptedAnswer":{"@type":"Answer","text":"No. The 905 figure counts verified systems retrieving two Google Ads update articles while building answers. It establishes access and consumption, but the log does not show whether an answer cited groas, recommended it, or produced any inquiry or revenue."}},{"@type":"Question","name":"How were the AI bot requests in the groas report verified as real?","acceptedAnswer":{"@type":"Answer","text":"The true client IP behind the CDN was matched against IP ranges published by the providers, such as OpenAI's published ranges for its bots. A User-Agent string alone is not proof, since anyone can fake a familiar crawler name. On the two update-article paths, verified requests made up 98.4% of recorded hits."}},{"@type":"Question","name":"Why should raw AI bot hit counts not be trusted?","acceptedAnswer":{"@type":"Answer","text":"A crawler name in a User-Agent is a claim, not identification. In the groas log, an RSS path recorded 545 total hits but only 3 verified requests and 0 live-answer fetches. A separate study across three sites found only 49.5% of 16,548 requests claiming known crawler names verified as genuine, with 39.3% spoofed."}},{"@type":"Question","name":"What evidence proves an AI assistant is actually recommending your business?","acceptedAnswer":{"@type":"Answer","text":"You need observed answers, not server logs: run a fixed prompt set on a schedule and record citations, mentions, recommendations, and share of voice. Even that is not enough for revenue reporting, which requires qualified inquiries, pipeline, and closed-won revenue tied to the activity."}},{"@type":"Question","name":"If a topic is missing from an AI visibility report, is that a content gap?","acceptedAnswer":{"@type":"Answer","text":"Not necessarily. The groas export was a truncated list of paths that registered hits, so a missing topic could have no demand, sit outside the exported paths, or simply not have been fetched during the 30-day window. Before accepting a gap diagnosis, ask for the full path export, the prompt set, the observed answers, and the measurement window."}},{"@type":"Question","name":"Do AI search summaries reduce clicks to websites?","acceptedAnswer":{"@type":"Answer","text":"In Pew's study of 68,879 searches, users clicked a traditional result 8% of the time when an AI summary appeared, versus 15% without one, and about 1% clicked a link inside the summary. These figures describe that study's behavior; they are not a conversion rate for any individual business or page."}},{"@type":"Question","name":"Should I create more content like my most-fetched pages to get more AI traffic?","acceptedAnswer":{"@type":"Answer","text":"Not automatically. The most-fetched groas pages were dated Google Ads update articles, which assistants retrieve for what-changed questions but which may not address buyer objections like who owns the outcome, speed, and cost. Check the gap against sales objections first, and fix broken rungs like access failures or citations without inquiries before chasing volume."}},{"@type":"Question","name":"What should I ask an AI search or GEO provider to report?","acceptedAnswer":{"@type":"Answer","text":"Ask for three separate reports: IP-verified fetches by page split into index crawls and live-answer fetches; recorded answers to a fixed prompt set with citations, mentions, recommendations, and share of voice; and qualified inquiries, pipeline, and closed-won revenue where measurable. Their claims cannot be blended, and a visibility score relabeled as ROI is not a substitute for business outcomes."}}]}]}
```
