A Haus Research audit failed 34.7 percent of Perplexity's number citations: Reveneau

Why it mattersFor a brand relying on generative answer engines, being cited is only half the job, because the number stated next to your name may not match your own site, and a reader who checks will end up on a different page.

Haus Research published a report on 2 September 2026, filed as HR-2026-09, measuring whether the pages Perplexity cites next to a number actually contain that number. Of 1,826 citations that Perplexity's search models attached to a sentence stating a figure, 34.7 percent pointed at a page that either would not open or did not contain a single figure from that sentence.

How the measurement was done

Haus Research picked 210 well-known technology companies by hand and asked 310 questions across ten templates: founding, latest funding round, headcount, entry price, headquarters, revenue, disclosed breaches, current CEO, acquisitions, and paid-tier uptime SLA. Every company got one question and 100 got a second on a different template. Questions were sent at temperature zero to perplexity/sonar and perplexity/sonar-pro, with GPT-4.1 and a web plugin used as a control.

For each answer the researchers pulled the specific numbers stated, money amounts, percentages, magnitudes, years and any run of three or more digits, and asked whether the cited page's visible text contained at least one of them. Values were normalised so that "$185 million", "$185M" and "185000000" all count as the same number.

Where the citations failed

The 34.7 percent failure rate breaks down in a way that matters for what a fix would look like. Dead links accounted for only 1.3 percent, so broken links are not the problem. Gated or paywalled pages the model could not read behind accounted for 16.1 percent. Another 2.5 percent were client-rendered pages that the retrieval crawler could not read. The largest remaining group, 16.1 percent of the readable pages, opened cleanly and simply did not contain the number Perplexity had stated next to them.

Two thirds of the 872 claims in the sample carried more than one citation. When Haus Research scored per claim instead of per citation, passing a claim if any of its citations contained the figure, the failure rate dropped to 14.4 percent. Perplexity is placing several markers on the same claim, which raises the pass rate but does not answer which one the reader is meant to trust.

A citation can be real and still fail

The GEO conversation has focused on getting cited. This report is a reminder that the citation is only half the job. If a page is behind a paywall, uses client-side rendering that the crawler cannot execute, or states a headline figure in an image or a PDF rather than the body text, the model may still cite it but the reader who clicks through reaches a page that does not confirm the number. The result looks like a real citation and behaves like a failed one.

Two practical readings for a product marketing team. First, put the numbers Perplexity would want to quote for you, revenue, headcount, funding, uptime, into the plain HTML body of the page the crawler can read. A PDF report, a rendered dashboard or a screenshot of a chart will not count. Second, if you audit brand visibility across generative engines, add a citation-content check: pull the pages the engine cites for your brand and confirm the numbers appear on them, because a good citation count with a bad content match is a false positive.

The report covers Perplexity's sonar family and does not measure ChatGPT, Google's AI Overviews, or Claude, and the 210-company sample is picked rather than randomised. Read the 34.7 percent as the size of the problem for one engine on one kind of question.

Source

Are Perplexity's citations quotable?, HR-2026-09, Haus Research, 2 September 2026.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.