We traced 391 web sources behind 10 AI answers to software pricing questions. Review aggregators appeared in 9 of 10 answers and were cited zero times. Here is the data and its limits.
We asked an AI assistant ten commercial pricing questions with web search enabled, then traced every checkable claim in its answers back to specific URLs. That produced 391 source pages across 149 domains.
The single most striking result: SoftwareAdvice appeared in the candidate sources for 9 of the 10 answers and was never cited once. Seventeen pages fetched, zero used. TrustRadius shows the same pattern across 25 pages, with a caveat covered below.
The overall citation rate in this corpus was 33%. These domains sit at zero.
Present is not the same as read
Every AI answer comes from somewhere, and there are two very different facts about any page in that pipeline:
- Present — the page was in the candidate set for the question.
- Cited — the model itself declared it used that page.
Most visibility tooling measures the first and reports it as though it were the second. They are not the same measurement, and this corpus shows them coming apart sharply.
| Metric | Value |
|---|---|
| Source pages traced | 391 |
| Unique URLs | 389 |
| Registrable domains | 149 |
| Pages cited | 129 (33.0%) |
| Per-question citation rate | 27%–46% |
That range matters. Every single question landed between 27% and 46%, so the measurement is stable — a domain at 0% across nine questions is not a sampling artifact.
Citations were spread, not concentrated
Before the aggregator result, a finding that surprised us in the other direction: there was no small group of dominant sources.
- 93 of 149 domains had at least one cited page
- It took 29 different domains to account for half of all citations
- The most-cited single domain reached only 6 pages (salesforce.com)
There is no citation oligarchy here. Answers were assembled from a wide, shallow spread of sources — mostly vendor documentation and smaller independent write-ups.
The aggregator result
| Domain | Questions present | Pages fetched | Times cited |
|---|---|---|---|
| softwareadvice.com | 9 of 10 | 17 | 0 |
| trustradius.com | 9 of 10 | 25 | 0 |
The obvious explanation is that these pages contain no prices, so a method tracing price claims could never connect them. We tested that, and for SoftwareAdvice it fails.
SoftwareAdvice pages are price-dense and substantial. Median 13 price figures per page. Only 2 of 17 pages had none. Median length on the main host is 25,186 characters — longer than the median page that did get cited (23,635).
| Group | Pages | Median price figures | Pages with none | Median length |
|---|---|---|---|---|
| Cited pages | 129 | 25.0 | 12% | 23,635 |
| Uncited, on domains cited elsewhere | 151 | 3.0 | 35.8% | 11,913 |
| TrustRadius + SoftwareAdvice | 42 | 9.5 | 21.4% | 23,325 |
These pages carry roughly a third of the price density of cited pages, but three times that of ordinary uncited pages, at comparable length.
Our own pipeline reads SoftwareAdvice fine
The second possible explanation is that our matching was quietly failing on these pages — prices trapped in tables or widgets that survive fetching as unusable text.
Matching requires a product name and a distinctive price token to appear together inside one passage. Against that gate:
| Domain | Pages clearing the gate |
|---|---|
| softwareadvice.com | 10 of 17 (58.8%) |
| Baseline, all uncited pages | 128 of 262 (48.9%) |
| trustradius.com (all subdomains) | 11 of 25 (44.0%) |
| — trustradius.com main host only | 7 of 13 (53.8%) |
| — solutions.trustradius.com | 4 of 12 (33.3%) |
SoftwareAdvice clears the gate at an above-average rate. Our pipeline extracts usable pricing claims from those pages more readily than from the average uncited page — and the model, given the same pages, used none of them.
TrustRadius is a weaker case and we are saying so. On a whole-domain basis it falls below the baseline, because roughly half its footprint here is a buyer-guide subdomain that is genuinely thinner: 5 of 12 of those pages had no price at all. Its main product and comparison pages behave like SoftwareAdvice; the family as a whole does not. The strong version of this claim applies to SoftwareAdvice.
The counter-example that makes this credible
YouTube looks like the same story and isn't.
YouTube accounted for 104 pages — 26.6% of the entire corpus — and was cited 4 times, a 3.8% rate. High presence, near-zero citation.
But YouTube transcripts had a median of 1 price figure, 47.1% had none at all, and median length was 8,853 characters, roughly a third of the aggregator pages. For YouTube, "not cited" is largely explained by "contains almost no prices." That is a fact about the content, not about the model.
Same surface pattern, different cause. We report it because a finding that survives only when you don't check comparable cases isn't a finding.
Limitations
These are substantial. Read them before quoting anything above.
- This is one model, not "AI engines." The corpus was built with Claude Haiku plus Tavily as a search layer. A third engine, GPT-4o-mini, was configured but returned quota errors on all ten questions and contributed nothing. Nothing here describes ChatGPT, Gemini, Perplexity, or any other system.
- Ten questions, two categories — 7 CRM, 3 project management. All business-software pricing.
- A 23-hour window. All ten traces were built between 5 and 6 August 2026. Retrieval behaviour changes; this is one narrow snapshot.
- Price claims only. This method traces numeric and product claims. It does not trace qualitative claims — three independent matching designs failed at that. A page cited for a non-price reason would not register here.
- "Cited" means the model declared it. A model can read a page without attributing it. We record declared citation and inferred influence as separate quantities, and this post uses the declared one.
- We cannot name model versions. The stored data records engine labels, not versioned identifiers.
A larger run across more categories and a working second model would establish whether any of this generalises. That work is in progress and the result will be published either way.
Method
Engines are asked with retrieval unforced — never with tool use required — because forcing retrieval would make "the engine searched" true by construction rather than by measurement. A trace is only built when at least two engines are usable and retrieval is confirmed.
Candidate sources come first from what the engine declared it cited, supplemented by search. Pages are fetched respecting robots.txt as a hard block, split into passages, and scored by embedding similarity only where a product name and a distinctive price token occur together in the same passage. The lexical requirement is the gate; the embedding decides the ranking.
Full method and limitations: geoscanai.co/trace/methodology
FAQ
Does this mean review sites don't matter? No. It means that in this corpus, for pricing questions, one model had them available and used other sources. Their value to human buyers and to classical search is a separate question this data does not address.
What did the model cite instead? Vendor documentation and smaller independent sites, spread thin — 29 domains to reach half of all citations.
Is ten questions enough? Enough to report a measurement with its limits, which is what this is. The per-question citation rate held between 27% and 46% throughout, so zero-of-42 across nine questions is not noise. Whether it generalises past CRM and project-management pricing, or past this one model, is untested.
How do you define "cited"? The model's own response declared it used that URL. That is narrower and more verifiable than inferred influence, and the two are stored separately.