AEDAnswer Engine Daily
← All stories

Perplexity’s Decisions API now searches the web before answering, but doesn’t say whether sites can see or block their pages being used as evidence

Sandeep Amar · November 10, 2026 · 6 min read
Perplexity’s Decisions API now searches the web before answering, but doesn’t say whether sites can see or block their pages being used as evidence
LLMs · Answer Engine Daily

Perplexity said on November 3, 2026, that its Decisions API, a developer product that answers questions with probabilities instead of text, can now search the web before it answers. On the 100 public questions of PWS-Decisions, a benchmark of questions that depend on current web facts, accuracy rose from 51% to 86% with the setting on, the company said in a blog post.

However, the 86% comes from Perplexity’s own run, the post’s text does not name the benchmark’s authors, and a second, larger jump in the same post comes from a test Perplexity designed itself. The post also does not say whether a website can see, or opt out of, its pages being used as evidence.

That matters because of where answers go. Perplexity’s examples include qualifying sales accounts (“Raised funding in the last 90 days?” across an entire account list), checking for a newer software release and confirming “Is this event still on?” before an agent books. Each answer, the post says, is “a probability a developer’s code can act on, with the web sources behind it.” The post does not say whether anyone beyond the developer sees those sources, or whether the feature is in beta or generally available.

What is a decision model?

A model that returns a number instead of written text. According to the post, a developer sends content as “state” (text, JSON or images) with one or more questions. A yes-or-no question returns the probability of yes; a choice question returns a probability for each option; a score question returns a probability for each level of a rubric and the expected score.

Without grounding, the model has two sources, Perplexity says: the state and what it learned in training. The company says that is enough for “Does this review report a defect?” but not for “Is this person still the CEO?”, since the answer may have changed after training. Web grounding adds a third source, the web, searched at the time of the request.

How does web grounding work?

Developers add one setting, “mode”: “web_grounding”. The post and Perplexity’s cookbook describe these steps:

  • Query building: For each question, the API gateway builds a Perplexity Fast Search query from the question text and “as much of your state as fits in 256 bytes,” the cookbook says.
  • Evidence: The gateway merges results, drops duplicate URLs and appends each page’s title, URL and text to the state, the post says.
  • URLs as state: A developer can pass a URL instead of text, and the page itself comes back as evidence, according to the post.
  • Visibility: The response adds a search_results field listing the pages the model read.

The 256-byte limit has a practical effect. In the cookbook’s example, asked whether Navi Pillay was awarded the 2026 Nobel Peace Prize, the ungrounded model returned 0.01 and the grounded one 1.00. With 600 characters of unrelated text placed before the claim, the grounded request returned 0.07, with five pages about office moves as its evidence. This means where a claim sits in a request decides which pages are fetched, and so which pages the answer rests on.

Neither page says how Fast Search ranks results.

How much did accuracy improve?

It depends on the test. The post gives two different before-and-after pairs:

  • 328-claim test: Perplexity said it sent 328 true-or-false claims, split between October 2026 news and rare Wikipedia facts, to the same model twice. Each true claim had a false twin differing in one detail. Accuracy rose from 56.4%, which the post calls “close to guessing”, to 97.3%, and errors fell from 143 to 9.
  • PWS-Decisions: Perplexity ran the 100 public questions with pplx-decider-v1.2 on November 4, 2026, a day after the post’s date, and got 86 right. The 51% without grounding is from the benchmark authors’ leaderboard, which Perplexity says matches its own run.

The cookbook adds a smaller run of 24 labelled claims (8 long-settled facts, 16 events from September and October 2026). Without grounding, no true recent claim scored above 0.24; with it, all 24 came back right.

The 328 claims are not linked in the post, and the leaderboard figures come from the benchmark authors, who have not been independently checked here. These results are Perplexity’s own account.

Are the speed figures comparable?

No. The post says “the tradeoff is speed”: in the 328-claim test, median latency rose from 139 ms to 366 ms, and 95th-percentile latency from 514 ms to 742 ms.

On PWS-Decisions, Perplexity measured a 370 ms median from its own client on the 100 public questions, against 560 ms that the benchmark authors report “for their API” on the full 300-question set, measured from their client on October 29. The post does not say which API that is. The post says that against the published leaderboard, the grounded API has “the highest public-set accuracy and a median response under 400 ms.”

What does it cost?

Each completed search costs $0.001 on top of input tokens, Perplexity said; the cookbook puts it as $1.00 per 1,000 searches. The search evidence itself counts toward input tokens. A request can carry up to 100 questions.

What this means for AEO

In my view, this is a quieter AEO surface than Perplexity’s answer engine, and a harder one to measure. Perplexity’s own examples (recent funding, a newer release, whether an event is still on) are factual checks about companies, products and events. Some page answers them, the answer becomes a probability, and someone else’s code acts on it. Nothing in the post suggests a click or credit goes back to the page, and the post points site owners to no report of it. Google and Bing at least share some AI search data now, as I wrote earlier.

The practical step is to state the facts these checks look for (release versions, event dates and status, leadership, funding) plainly, near the top of the page and in the initial HTML, since the post does not say how page text is extracted. If you build on the API, log search_results. It is the only window into which pages decide the answer.

Questions the post leaves open:

  • Opt-out: Can a website see that its page was used as evidence, or stop it?
  • Extraction: Does Fast Search read rendered pages or only the initial HTML?
  • Evidence volume: How many results per question are appended, and how are they ranked?
  • Benchmark: Who built PWS-Decisions, and can its leaderboard be checked independently?
  • Test set: Will Perplexity publish the 328 claims so the 97.3% can be checked?

Sources

More in LLMs