Perplexity Decider: You can download the bouncer, not the house rules

You have probably seen the pitch. An agency shows you a "Perplexity score" for your pages. The threshold is 0.7, they say, and below that you never get cited. Where does the number come from? Reverse engineering. In other words, a guess with a spreadsheet attached.

Since November 12 there is an actual data point. And it is more interesting than any spreadsheet.

What's new

• Decider v1.3 and Decider Flash v1.3: two decision models from Perplexity Research. The large one is built on Qwen3.8-27B; the small one is distilled into Qwen3.5-4B.
• What they do: they answer yes/no, multiple-choice and score questions and return a probability with each answer. Text and images go in; no free-form prose comes out.
• Open: the weights are on Hugging Face under an Apache license, with access through OpenRouter and the Perplexity Decisions API.
• The line marketers should care about: Perplexity says it can use Decider to judge the quality of pages as it crawls them. Another head decides whether a user query triggers a web search at all. Reranking search results by relevance is listed as a use case too.

The benchmarks (for example 91.6% accuracy across 63 query-classification tasks) are Perplexity's own numbers. You may know this model class already: Google launched the Gemini Decisions API a day earlier. The difference is that Perplexity hands over the model itself.

You get the bouncer, not the house rules

What I find genuinely interesting: for the first time you can run the same kind of gatekeeper an AI search engine uses on your own machine. Until now GEO has been flying blind. Change a page, wait weeks, hope.

But. A bouncer works from house rules, and those are missing. Which exact question Perplexity asks at crawl time, which rubric sits behind it and what score gets a page dropped are not in the post. Neither is whether production runs these exact weights. Perplexity itself advises tuning decision thresholds on your own labeled data.

So anyone now selling you "the real Perplexity score" is still selling an assumption. Just with a better tool.

The real wake-up call sits one step earlier

For me the key sentence is not the one about page quality. It is the head that decides whether to search at all. If the AI answers from the model, there is no source to cite. However good your page is.

That is the agent logic I keep coming back to: a model decides what your customer gets to see. Now you can see it spelled out as architecture. First the switch: search or not. Then the bouncer: is the page good enough. Then the ranking. Three yes/no questions before your brand name ever comes up.

And AI does not fix a thin page. A quality filter loves clear facts, direct answers and clean structure. That is the unglamorous groundwork, not a hack you bolt on afterwards.

What I would do now

• Write your own house rules: five yes/no questions every page has to pass (Does it answer the customer's question in the first paragraph? Does it show price and an as-of date? Is it clear who stands behind it?).
• Run those questions with a decision model against your 20 most important pages, Decider via OpenRouter or locally. What matters is how your pages compare with each other, not the absolute score.
• In parallel, ask several AI systems five real customer questions without your brand name, and log the date. That shows you whether they search and cite at all.

Bottom linePerplexity just gave you the bouncer. You still have to write the house rules yourself, and make them stricter than his.