Perplexity Brand Monitoring: How to Track What Perplexity Says About Your Brand

Perplexity is a search-first answer engine. When someone asks about your brand, it retrieves web pages, passes them to a language model, and shows the answer with numbered citations.

Three mechanics matter for monitoring:

• Whole-source citations. Perplexity cites entire sources by integer index. Anthropic's Claude, by contrast, cites short text spans. So a Perplexity answer tells you which pages it leaned on, not which sentence on each page.
• Model routing. As of September 2026, secondary reports describe the Free tier as using "Best", an auto-router, while Pro and Max users pick from a menu of about eight models, including Sonar 2, GPT-5.6, Gemini 3.7 Flash and Claude Sonnet 5. Perplexity does not publish one model that "is" the app.
• Retrieved is not the same as cited. Engines that expose both lists show a larger retrieved set than cited set. Mechanistic research on how models cite (arXiv:2606.28358) found citation output can be changed without changing the sources, which means a citation list is a noisy proxy for influence.

What this means: the same question can produce different wording and different sources depending on the model and the day. Your log needs to capture those conditions, or the numbers will not compare.

It helps to split an answer into four stages.

You fully control access. You control the facts printed on your own pages. You can influence, but not control, whether third-party sites that Perplexity cites describe you correctly.

You do not control retrieval ranking or model phrasing. A critical survey of 45 GEO studies (arXiv:2607.14035, July 2026) found no technique with a stable, cross-platform causal effect on discoverability. Treat any promise of a guaranteed Perplexity citation as unsupported. The work that holds up is removing blocks and fixing wrong facts.

Monitoring answers is pointless if Perplexity cannot read your pages. Check access before anything else.

robots.txt

Perplexity documents PerplexityBot as its search-indexing crawler and says it is not used to crawl content for AI foundation models. Blocking it keeps your pages out of Perplexity's index. A minimal allow rule looks like this:

User-agent: PerplexityBot
Allow: /

Check for blanket rules such as User-agent: * with Disallow: / that catch it by accident. Note that Perplexity-User, the fetcher triggered by a user's request, "generally ignores robots.txt rules" per Perplexity's docs. Blocking it there is not reliable.

Firewall and CDN

A firewall or CDN can block a crawler that robots.txt allows. The result is often a 403 or a challenge page, and no dashboard tells you. Cloudflare changed its AI crawler defaults on 2026-09-15 for new domains, new sites on existing accounts and existing free-tier customers, so a site can lose AI access without anyone editing it. Request a page with the PerplexityBot user agent and confirm you get a 200 with real content. 99Visibility's AI Crawler Checker runs that request for each AI crawler and shows which ones get turned away.

JavaScript rendering

Independent testing re-confirmed in June 2026 that PerplexityBot does not execute JavaScript. Vercel's crawler study reached the same conclusion for ChatGPT and Claude crawlers. If your pricing, product names or company facts only appear after scripts run, Perplexity may never see them. Compare the raw HTML with the rendered page and move missing facts into the server response.

Once access is clean, build a repeatable log. The goal is to connect each claim Perplexity makes about you to the source it cited.

1. Write a fixed question set. Pick 10 to 30 buyer questions: "What does [brand] cost?", "[brand] vs [competitor]", "Is [brand] good for [use case]?". Keep the wording identical across runs.
2. Fix the conditions. Use a logged-out session or a clean account. Record the plan tier and model setting ("Best" or a named model). Note the date.
3. Capture the full answer. Save the answer text and every numbered citation with its URL.
4. Split the answer into claims. One row per factual claim: price, founding year, features, headquarters, competitor comparison.
5. Map each claim to its cited source. Note which numbered source sits next to the claim. Mark the claim correct, wrong, or unverifiable against your own records.
6. Tally sources across runs. Count how often each domain is cited for your brand questions. Recurring third-party pages are the ones shaping your description.
7. Repeat on a schedule. Monthly is enough for most brands. Run again after any site change or correction.

Score wrong facts as a rate over answers returned, not as raw penalty points, so a bigger question set does not punish you. The guide on measuring brand accuracy in AI answers covers the scoring choices. For a single snapshot before you build the log, try the free check of what Perplexity says about you.

You can monitor Perplexity through the consumer app or through its API. They are not the same instrument.

The Agent API returns a search_results list with fields such as url, title, snippet, date and last_updated, plus integer citation indexes. That makes source logging easy to automate. Two cautions apply:

• Web search is opt-in on the Agent API. Perplexity's migration docs say so. Code still pointed at the old Sonar Chat Completions endpoint, which retired on 2026-09-27, can return answers with no web grounding at all. An ungrounded answer about your brand measures the model's memory, not Perplexity search.
• The API model is not the app. No single API model matches the Free tier's auto-router or the Pro model menu. Sonar is the closest pin, as of September 2026.

Use the API for scale and the app for spot checks of what real users see. Label every row with which one produced it. The comparison of API answers and consumer app answers explains the gaps in more detail.

Your log will show two kinds of problems. Each has a different fix.

Wrong facts that trace to your own site

If Perplexity cites your pricing page and still gets the price wrong, check the page. Look for contradictions between the pricing page, about page, footer and help docs. State the fact once, plainly, in HTML that loads without JavaScript. A specific stated fact (a price, a number, a one-line definition) on a fetchable page gives the engine a sentence it can quote. That is the only positive experimental signal for on-page content in the evidence. It is not a guarantee.

Wrong facts that trace to third-party sources

If the wrong claim sits next to a review site, directory or old article, the fix is upstream. Contact the publisher with the correct fact and a source. In the evidence, being named accurately on the sites engines cite is associated with appearing in answers. Treat it as an association, not a promised effect.

Track each error from first sighting to resolution. An AI error log and correction tracker keeps that history so you can see which fixes stuck.

Hostile content

Some third-party pages contain hidden text or injected instructions aimed at AI systems. Research on search agents (arXiv:2606.16821) shows manipulated content in a result set can change what an agent endorses. If a cited source looks manipulated, record it as a reputational risk. Never use the same tactic yourself.

• Skipping the access check. Teams spend weeks logging answers while a firewall rule returns 403 to PerplexityBot.
• Ignoring model routing. Comparing a "Best" answer from March with a Claude Sonnet 5 answer from October tells you nothing about trend.
• Treating citation as use. A page that is not cited may still have been read. A cited page may have contributed little.
• Blending engines into one score. Otterly's 2026 experiment saw Perplexity, Claude and especially Copilot cite the test sites while Google's AI surfaces almost never did. A blended number would hide that split. See whether one dashboard can track every AI engine before trusting a single figure.
• Changing the questions every run. New wording means new results. Hold the set fixed and add questions as a separate cohort.
• Chasing tactics without evidence. Schema, llms.txt and "AI-friendly" rewrites have no credible evidence behind them for AI answers. Spend the time on access and facts.
• Running untagged API calls. If you cannot tell whether web search was on, you cannot tell what you measured.