Why visibility in one AI engine says little about the next: one month of retrieval-to-citation data from eight B2B projects
A comfortable assumption runs through a lot of GEO work: optimize a page once, and it shows up everywhere. In ChatGPT, in Perplexity, in Google’s AI Overview. One effort, three channels served. I measured that assumption for a month across eight live B2B projects. It does not hold.
Across the projects, the three engines cite almost entirely different sources. Two of them cite, on average, only one in five of the domains that any other engine also puts in a citation. Work done for ChatGPT is barely work done for Perplexity. That changes how you should measure, budget and report AI visibility. The numbers behind it are below, along with the tables they come from.
Every AI answer happens in two steps
Ask ChatGPT, Perplexity or Google for a recommendation, and more happens behind the scenes than a block of text. First the system gathers a set of pages it judges relevant to the question. That is retrieval. From that stack it then selects which sources actually feed the written answer and appear beneath it as citations. That is citation.
Here is the image I will keep using. Being retrieved means being let into the anteroom. The model considers your page relevant enough to look at. Being cited means being called into the room and named, where the customer reads the answer. Between the anteroom and the room sits a selection step. Almost nobody measures that step.
In an earlier piece I called this gap, retrieved yet not cited, the most valuable one in your AI visibility (the article is here). Back then with illustrative figures. Now with the numbers from eight projects. And with an observation I could not prove at the time: there is not one anteroom. There are three. Each engine runs its own, with its own doorkeeper and its own taste.
A word on the data before the numbers. Eight anonymized B2B projects, one month from 18 June to 17 July 2026, tracking a fixed set of commercial prompts daily through the Peec.ai platform. Together more than 13,000 retrieved source domains, over 5,000 of them distinct, spread across roughly 17,000 domain-engine observations. Three sector groups: manufacturing, tourism and rail, other B2B. One project also tracks Gemini, but on a base too small to compare, so I keep that channel out of every comparison. All figures are computed at the domain level: of the sources an engine retrieved, the share that made it into at least one answer. I call that the citation rate.
The funnel leaks, and the engine decides how badly
Pooled across all eight projects, 55.6 percent of retrieved domains received at least one citation. A little over half. Even in the most permissive project, a third of the retrieved sources vanished without a trace; in the strictest, more than half. Being retrieved is a precondition, not a promise.
It gets interesting once you split the citation rate by engine. Averaged over the projects, it runs at 76.8 percent for Google AI Overview, 47.7 percent for Perplexity and 40.9 percent for ChatGPT. This is no artifact of pooling. Google AI Overview stays between 68 and 82 percent in every single project, and ChatGPT is the stingiest of the three engines in six of the eight. In the two exceptions, Perplexity cites slightly more sparingly still.
| Project | Sector | AI Overview | ChatGPT | Perplexity |
|---|---|---|---|---|
| P1 | other B2B | 74.2 | 41.8 | 62.6 |
| P2 | other B2B | 80.6 | 30.3 | 41.5 |
| P3 | manufacturing | 82.0 | 33.8 | 46.1 |
| P4 | tourism/rail | 76.5 | 32.1 | 37.8 |
| P5 | manufacturing | 78.3 | 64.2 | 57.9 |
| P6 | other B2B | 74.7 | 33.1 | 50.7 |
| P7 | manufacturing | 79.2 | 43.1 | 47.2 |
| P8 | tourism/rail | 68.6 | 48.7 | 37.7 |
| Mean | 76.8 | 40.9 | 47.7 |
Table 1: citation rate in percent (cited domains divided by retrieved domains), by project and engine.
The reading is uncomfortably clear. Google AI Overview cites almost everything it retrieves. That fits its architecture: retrieval here is classic Google ranking, and Google says itself that its AI features draw on the same index and the same ranking systems as ordinary search (Google Search Central). What survives that hurdle usually carries through to the answer. ChatGPT retrieves broadly, in four of the eight projects more widely than any engine, and cites the most sparingly. The doorkeeper in the anteroom is strict. For Google AI Overview, retrieval is nearly a citation already. For ChatGPT it is roughly a two-in-five chance.
Keep this sentence for your reporting: “we were retrieved” means something completely different depending on the engine.
Three engines, three near-separate source pools
If the engines drew on a shared canon of trusted sources, the sets of domains they cite should overlap heavily. They do not. For every engine pair in every project I measured how much the sets of cited domains overlap. The value stays between 0.12 and just over 0.21 in every project. Put plainly: of five domains one engine cites, on average only one is also cited by another engine, for the same tracked prompts, in the same month.
| Project | AIO × ChatGPT | AIO × Perplexity | ChatGPT × Perplexity |
|---|---|---|---|
| P1 | 0.133 | 0.175 | 0.179 |
| P2 | 0.170 | 0.200 | 0.214 |
| P3 | 0.128 | 0.200 | 0.154 |
| P4 | 0.155 | 0.195 | 0.163 |
| P5 | 0.181 | 0.214 | 0.149 |
| P6 | 0.125 | 0.133 | 0.173 |
| P7 | 0.186 | 0.140 | 0.198 |
| P8 | 0.161 | 0.194 | 0.166 |
Table 2: overlap of cited domains per engine pair, measured with the Jaccard index. 0 means no shared domain, 1 means identical sets. The method note at the end explains the metric.
This is the most consequential finding in the analysis, and the reason for the title. An overlap of 0.12 to 0.21, stable across eight projects and several sectors, is hard to square with the common working assumption that one well-optimized source will carry across engines. Occasionally it does. Four times out of five it does not, at least within a month of tracked commercial prompts. “AI visibility,” then, is not one target. On this evidence it is three loosely coupled targets, and you have to aim at, budget for and report each of them on its own.
Which source type gets cited depends on the engine
That leaves the small-scale question. Does an engine favor particular kinds of source, editorial outlets say, official corporate pages, institutional senders? It does, but the direction flips from engine to engine, and the effect is small.
| Source type | ChatGPT | Perplexity | AI Overview |
|---|---|---|---|
| Corporate | 44.5 | 49.1 | 76.9 |
| Institutional | 42.6 | 41.2 | 79.4 |
| Reference | 36.7 | 45.4 | 79.8 |
| UGC | 33.5 | 41.1 | 82.1 |
| Editorial | 32.8 | 50.4 | 75.6 |
| Other | 33.5 | 29.6 | 71.7 |
Table 3: citation rate in percent by source type and engine, pooled across projects. Tracked competitors and the project’s own domain are excluded, because they are cited almost by construction and would distort the comparison.
Look at the “Editorial” row. In ChatGPT, editorial domains convert worst of all types, at 32.8 percent. In Perplexity the same editorial domains sit at the top, at 50.4 percent. Same source type, opposite result. In Google AI Overview, meanwhile, every type lands between roughly 72 and 82 percent, and the difference is not statistically meaningful. The engine cites across all types.
I draw two conclusions. There is no universally citable source type, and the clearest proof is the editorial mention, bottom in ChatGPT and top in Perplexity. And even where the type effect is statistically clean, it stays small next to the engine effect (the effect sizes are in the method note). What kind of site a source is matters far less than which engine it is trying to appear in. Advice of the form “engines love editorial coverage” or “engines prefer official corporate pages” is, in this data, true for exactly one engine and false for another.
What this means for your measurement and your work
Four consequences follow from the three findings, straight from the numbers, with no detour through generic advice.
1. Keep two metrics per engine, not one visibility score. Retrieval rate and citation rate belong side by side, per engine. The gap between them is itself the diagnosis. A blended “visibility” hides exactly the gap the work targets.
2. Locate the failure before you fix it. When your page is missing from answers, first establish where it fails. A domain that is never retrieved has a technical or an authority problem, and the answer is a new page or signals from outside. A domain that is retrieved but not cited has a selection problem, and the answer is to make the existing page more citable. Two different sites of work, two different tools.
3. Do not count on transfer. Presence in one engine tells you little about the next on this data. Set goals per engine and check per engine, rather than maintaining a single number for “the AI.”
4. Weight the effort by the engine’s mechanics. For Google AI Overview the fight is almost entirely upstream, at retrieval: what survives Google ranking usually gets cited. For ChatGPT the fight continues after retrieval, at selection. How ChatGPT makes that selection I have taken apart elsewhere (how ChatGPT chooses its sources).
To take away, a template you can paste into a spreadsheet and fill per engine. First the metrics, then a diagnostic rule that turns the two rates into the next action.
AI visibility, metric split per engine (per prompt set, monthly)
Engine | retrieved | cited | citation | typical benchmark
| domains | domains | rate | (this dataset)
--------------+-----------+----------+----------+-------------------
AI Overview | | | | ~77 %
ChatGPT | | | | ~41 %
Perplexity | | | | ~48 %
Diagnosis per prompt and engine:
- not retrieved -> technical problem or authority gap
(new page, third-party signals)
- retrieved, not cited -> selection problem
(make the existing page citable)
- cited -> monitor, hold the positionThe benchmarks are not targets. They are the frame of reference from this dataset. If your ChatGPT citation rate sits at 20 percent while the reference is around 41, you know where to start. If your AI Overview rate is 40 where 77 is usual, your problem is almost certainly retrieval, the classic ranking step. The wording barely comes into it.
When these numbers hold, and when they do not
I am not selling a constant of nature. Five limits are worth knowing, or you will overstretch the claim.
First, and most important: “retrieved” here is an event inside the engine’s answer process, as the monitoring platform logs it. It is not a hit in your server log. What is measured is the path from retrieval to citation inside the answer pipeline. The path from crawl to citation is something else. Do not conflate the two. Second, the window covers a single month, with no control for seasonality or model updates. The figures carry a date. Third, the source-type classification is heuristic and was not hand-audited. Fourth, the prompt sets are individual clients’ commercial campaigns, not a neutral sample of the web; the numbers describe the source ecosystems these engines draw on for these commercial topics. Fifth, retrieval and citation counts come from the same measurement stack, with no external check.
Every relationship here is correlational. Nothing explains why an engine cites or drops a retrieved source. It shows how often, and how differently across engines. What convinces me about the pattern is its repetition: an overlap of 0.12 to 0.21 across eight projects and several sectors is unlikely to be one market’s quirk.
Frequently asked questions
Is “retrieved” the same as “an AI bot visited my server”? No, and the confusion is expensive. “Retrieved” here means a URL from your domain entered the engine’s context during the answer process, as logged by the monitoring platform. A hit in your server log is a different event. This analysis measures the step from retrieval to citation inside the answer.
Does a 77 percent citation rate mean I can neglect Google AI Overview? The opposite. The high rate only means the selection after retrieval barely filters anymore. The fight sits before it: AI Overview retrieves what survives classic Google ranking. That is where the hurdle sits, at retrieval itself.
If I am visible in ChatGPT, am I also visible in Perplexity? On this data, barely. The cited source pools of any two engines overlap only at 0.12 to 0.21. Four of five cited domains are cited by one engine only. Plan per engine.
Which source type is most likely to be cited? It depends on the engine. Editorial domains rank top in Perplexity (50.4 percent) and bottom in ChatGPT (32.8 percent). In Google AI Overview the engine cites across all types at roughly 72 to 82 percent. There is no universally citable type in this data.
Is an llms.txt enough to make engines favor my page? It does not look that way. If a large share of visibility is decided at selection, and that selection works differently in every engine, the leverage lies in understanding retrieval and selection per engine. A file that declares intent moves little. Why the file does so little I have taken apart in the llms.txt guide.
In closing
It started with the comfortable assumption that well-optimized content carries across all AI answers. Behind it stand, in fact, three anterooms with three doorkeepers who, faced with the same questions, agree on almost different sources. Test it on your own case. Take five prompts where you want to be named, and measure retrieval and citation separately, per engine. The engine where you are retrieved often and cited rarely is your fastest lever. There, the existing page already stands in the anteroom. It only has to get into the room.