33% of Perplexity Citations Are Wrong: How to Be a Source AI Quotes Correctly

As of September 2026, two independent audits published within the same week told one story. Haus Research sampled Perplexity's answer citations and found that about one in three — 33% — did not contain the specific number the AI had attributed to them. Days earlier, an investigation documented three websites that produced 215,128 "best software" pages built for AI answer engines to read, and Perplexity cited them anyway. Both write-ups reached the Hacker News front page inside 48 hours, at 86 and 197 points.

Put those together and you have the state of AI search for a business owner: answer engines quote sources loosely, and other operators are manufacturing sources to take advantage of that. You cannot fix either one. What you control is whether your own pages are the kind a machine can quote correctly — and whether the pages naming your competitors would survive a fact-check.

Why this matters for a local business

When someone asks Perplexity or ChatGPT for the "best HVAC company in Tampa" or the "best personal injury lawyer near me," the model does not think. It retrieves a handful of web pages, pulls sentences out of them, stitches those into an answer, and attaches citations. If your service area, your pricing model, your years in business, and your differentiators are not written in plain text on a page — with a number and a date — there is nothing for the model to lift. If a competitor has all of that on one clean page, the model lifts theirs and names them.

The 215,128-page investigation shows the tempting shortcut: pay to be listed on a site that mass-produces "top 10" pages tuned for AI. It works until it does not. Those networks get identified, filtered, and de-cited in waves — we covered one such takedown in our breakdown of the three sites behind 215,128 fake "best software" pages. Building your own quotable pages is slower and it compounds. Renting someone else's is faster and it evaporates.

Here is what quotable looks like in practice. A roofing company in Sarasota rewrites its homepage line from "the area's trusted roofer" to "roof replacements and storm repair across Sarasota and Manatee counties since 2009, licensed CCC1331745." Now an answer engine has a location, a service list, a start year, and a license number it can quote verbatim. The vague version gave it nothing. That is the difference between being read and being ignored, and it costs one afternoon of editing per page.

This is also the deciding factor if you are weighing whether to build your own agent stack for visibility work or bring in a firm. The audit-and-publish loop is not hard; it is relentless. A monthly cycle across five answer engines, dozens of customer questions, and every competitor page is a standing job, not a one-time project. Owners who try to run it between service calls do it once and stop.

The playbook

Six steps, in order. The first two are free and take an afternoon; the rest are a standing habit.

1. Audit what AI already says about your category. Open ChatGPT, Claude, and Perplexity. Ask each the three or four questions a real customer would ask — "best [category] in [city]," "[category] near me that does [specific service]," "how much does [service] cost in [city]." Write down every business named and every source cited. This is your baseline. Most owners have never done it once.
2. Check whether those citations hold up. For each source Perplexity cites, open it and confirm the claim is actually on the page. Given the 33% miss rate, a meaningful share of what ranks in your category is quoting sources that do not say what the AI claims. Where a competitor is named on the strength of a shaky citation, a correct, specific page from you can displace it.
3. Make every claim on your site verifiable. Replace "years of experience" with "serving Orange County since 2011." Replace "affordable" with "flat-rate diagnostic, $89 as of September 2026." Replace "fast response" with "same-day service on calls booked before 2 PM." A number, a place, or a date on every claim. Machines quote specifics and skip adjectives.
4. Publish one reference page per topic a customer researches. Pricing model. Service area with named neighborhoods or counties. Your process, step by step. What you do that the other guy does not. One page each, structured with real headings, updated with a visible "last updated" date. These are the pages answer engines pull from.
5. Earn mentions on sources AI trusts. That means real directories, review platforms with actual verified reviews, local news, trade association listings, and chamber pages — not listicle farms. Three solid third-party mentions that a fact-check survives beat 30 that get filtered out in the next cleanup.
6. Re-audit monthly and keep a log. Run the same questions from step 1 on the first of every month. Record which businesses are named, which sources are cited, and what changed. AI answers shift as models reindex; without a log you cannot tell progress from noise.

Citation-readiness checklist

Run this against your own site before you spend a dollar on visibility work.

• ☐ Every headline claim has a number, a place, or a date — no naked adjectives
• ☐ Pricing approach is stated in plain text, not hidden behind a "contact us" form
• ☐ Service area names specific cities, counties, or neighborhoods
• ☐ Each key page shows a visible "last updated" date within the last 90 days
• ☐ Your founding year and team size appear as text on the About page
• ☐ At least three third-party sources name your business and would pass a fact-check
• ☐ You have run the AI-answer audit at least once and saved the results

Common pitfalls

• Buying placement on ranking farms. The 215,128-page network is one of many. Being cited through them is a liability that reverses the first time the answer engine filters the source.
• Vague claims. "Trusted," "leading," "premier" — a model cannot quote these, and they signal nothing. Every one is a missed chance to be specific.
• Stale pages. Answer engines quote the last version they indexed. A price you changed in March that still reads January's number will be repeated back to customers wrong.
• Chasing the ranking, ignoring the citation. Getting named once is not the goal. Being named because a page that survives scrutiny says so is the goal.
• No measurement. If you are not logging monthly, you are guessing. The 33% figure exists because someone checked; you need the same discipline on your own category.

Start with the audit

You cannot fix what AI says about your category until you know what it says today and which sources it trusts to say it. That is exactly what our free AI Visibility Report delivers in 24 hours: the questions customers ask, the businesses named, the sources cited, and where your pages fall short of being quotable. From there the monthly loop is straightforward to run — the hard part is the first honest look.