Do the cited sources actually explain the recommendation? 128 checked one by one
We read the pages the AI says it consulted and looked for the recommended brand in them. 128 brands out of 128, with a median of six pages citing each one.
In this article
Short answer: yes. Of 128 companies recommended by Perplexity, all 128 appear in at least one of the pages the engine says it consulted. The median is six citing pages out of twelve sources examined. The citations are not decoration: they genuinely carry the answer.
This article exists because yesterday's ended on a caveat we had written ourselves: « we measure what is read, not what convinced ». A cited page is not necessarily the one that put a name on the list. Leaving your own caveat to sleep turns it into a stylistic flourish. It can be settled, so we settled it.
If the sources did not explain the recommendations, everything we have published for three days about where to get cited would be advice with no foundation. That is the kind of check you had better run on yourself.
What we measured, exactly
We start from the answers collected on 24 August - five global sectors, no country named anywhere. For each one we have two things: the text, hence the recommended companies, and the cited URLs. All that is left is to read those pages and see whether the recommended name is in them.
The trap decides everything, and it is brutal. A page we fail to read - access refused, timeout, content built in JavaScript - would make a brand look « unsupported » when it may well be supported. We would then be manufacturing the article's most spectacular result out of our own collection failures. So we classified into three states, never two:
| State | Meaning |
|---|---|
| Supported | the brand appears in at least one cited page |
| Unsupported | every cited page was read, the brand is not in them |
| Undetermined | at least one cited page stayed unreadable |
The published rate is computed on settled cases only, and the undetermined share is published next to it: 254 of the 300 cited pages could be read, or 85 %. We announce ourselves as a robot and do not impersonate a browser - a site that refuses robots is entitled to refuse us, and that counts as « unreadable », not as a fact.
The result
| Verdict | Count |
|---|---|
| Supported by at least one cited page | 128 |
| Unsupported | 0 |
| Undetermined (unreadable page) | 6 |
| Rate on settled cases | 100 % |
And the support is not thin. The median recommended brand appears in six of the twelve pages examined, with a mean of 5.9. This is not a coincidence of vocabulary: it is convergence.
The nine per cent that rest on a single page
11 brands out of 128 are supported by exactly ONE cited page. It is the most actionable number in the measurement, and it reads both ways.
As leverage: in those cases, a single page was enough to put a name into the answer. One national comparison site, one trade-press article, one post on a local platform - one page, and the company exists for the assistant. That makes a source strategy concrete rather than incantatory.
As fragility: if you are cited by a single page, your presence rests on it. That page changes its ranking, disappears, or stops being read - and you vanish from the answer without anything having changed on your side.
What this result licenses, and what it does not
It licenses saying that the source list is a real lever. Since recommendations are genuinely carried by what the engine read, and since - that was yesterday's measurement - those pages are national, few, and identifiable per sector and per language, then working to appear on them acts on the cause rather than on a symptom.
It does not license saying that appearing on a page is enough to be recommended. We showed that every recommendation rests on a source; we did not show that every source produces a recommendation. Cited pages mention far more companies than the engine keeps. Being cited is a condition, not a pass.
Every recommendation rests on a page that was read. Not every page that was read produces a recommendation. Conflating the two would be selling a dream.
A defect found on our side, not theirs
The same measurement surfaced a problem in our tool. Our name extractor returned 200 entries; 66, a third of them, were not company names - ranking labels glued to their citation markers, such as `Best overall: HubSpot CRM[2][3][4][6]`, or advice openers (« For most people »).
So we set those 66 aside before computing, which is why this article speaks of 128 brands and not 200. A detail that matters: the only five « unsupported » cases in the raw measurement were all artefacts of that kind - not one was an ungrounded recommendation. The fix is queued. We say so here because a number whose cleaning is hidden is not a number.
The limits
One engine.
Perplexity is the only one of the three that returns usable URLs. Gemini hides its own behind redirects, and ChatGPT queried through the API does not search at all. This result holds for an engine that searches, not for « AI » in general.
Five languages.
English, French, Portuguese, Spanish, German - the ones where our name extractor is verified. It still returns section headings in Japanese, Korean, Polish and Arabic; measuring citation accuracy with an extractor that gets the name wrong would produce noise presented as a result.
A single day
, and assistants are not deterministic: two runs of the same English prompt overlap only 62-76 %.
Presence is not causation.
The name is in the cited page - that does not prove this page is what decided. We narrowed yesterday's question, we did not dissolve it.
We promise nobody a place in ChatGPT's answers - nobody can do that honestly. What can be measured is where you stand, per language and per engine, and which pages that standing rests on.