AI Search

Perplexity's Decisions API Can Now Check the Web Before It Answers

Perplexity's Decisions API can now check the web before it answers, through a new web grounding mode that developers switch on by adding "mode": "web_grounding" to a request. In Perplexity's own test of 328 claims, accuracy jumped from 56.4% without grounding to 97.3% with it, and errors fell from 143 to 9. The improvement is dramatic, and it makes a simple point: a model that guesses from memory will get facts wrong, and a model that looks them up will mostly get them right. The cost is speed — the typical reply slowed from 139 milliseconds to 366.

What the Decisions API does

The Decisions API is not a chatbot. It returns probabilities instead of prose: yes or no, a multiple-choice pick, or a scored answer to a structured question. That makes it a building block for applications that need a machine to make a judgement — a prediction market, a fact-checking pipeline, a research tool that scores claims, a trading system that weighs evidence.

The problem with that kind of API is that a language model's knowledge has a cut-off date, and it is confident about things it has never actually seen. Ask it whether a recent event happened, or whether an obscure fact is true, and it will answer with a probability that reflects its training data, not the world as it is now. That is why the ungrounded accuracy in Perplexity's test was only 56.4% — barely better than a coin flip on the kind of claims the test used.

Web grounding changes the order of operations. Before answering, the model searches live pages, reads what it finds, and then assigns its probabilities based on evidence it has just retrieved. The grounded model is called pplx-decider-v1.2, and it was run on the benchmark on November 4, 2026. Developers enable the mode with a single field in the request, which keeps the change easy to adopt for anyone already using the API.

The numbers, and what they actually test

Perplexity's test used 328 claims drawn from October 2026 news and obscure Wikipedia facts. Each true statement was paired with a false copy altered in a single detail — the kind of test that catches a model which answers from a vague memory rather than from evidence. Without grounding, the API got 56.4% right and made 143 errors. With grounding, it got 97.3% right and made 9 errors.

On the PWS-Decisions public benchmark, accuracy went from 51% to 86%. That is a smaller jump than the internal test, but it is on a public benchmark rather than Perplexity's own, which makes it the more useful number for anyone comparing systems. The gap between the two results is itself informative: the internal test was designed around the failure mode grounding fixes, so it shows the best case, while the public benchmark shows something closer to a general case.

All of these figures are Perplexity's own, and no independent run has been reported. That does not make them wrong, but it does mean they should be read as a vendor's demonstration rather than a neutral measurement. The direction is plausible — retrieval helps factual accuracy, and every serious study of the technique says so — but the exact size of the gain will depend on the questions a real application asks, not the ones in Perplexity's test set.

Speed, cost, and the trade-off

Grounding is not free. The typical reply slowed from 139 milliseconds to 366 milliseconds, because the model now has to search, fetch and read before it answers. For a prediction market or a research tool, 366 milliseconds is still fast. For a high-frequency application that makes thousands of decisions a second, it may be too slow, and the ungrounded mode may still be the right choice for questions that do not depend on recent facts.

The price is small: each search costs $0.001 on top of the input tokens. At that price, grounding is affordable for almost any application that needs it, and the decision is about latency and design rather than budget. A developer can ground the questions where freshness matters — recent news, current figures, obscure facts — and skip it for questions where the model's training data is reliable.

Why it matters for SEOs and link builders

For SEOs, web grounding is good news and a warning at the same time. It is good news because an AI system that checks the web before it answers is an AI system that rewards accurate, well-sourced, up-to-date pages. The pages it retrieves are the pages it cites, and the pages it cites are the ones that shape the answer. A page that states its facts clearly, with dates and sources, is exactly the kind of page a grounding search finds useful.

It is a warning because the same system will find the errors. A page with an outdated figure, a misattributed quote, or a claim that was true in 2024 and false in 2026 will now be retrieved, read, and used as evidence — or, if the model is doing its job, retrieved, read, and discounted. Accuracy is no longer just a matter of reader trust. It is a matter of whether the machines that answer questions on your behalf will treat your page as a source or as a cautionary example.

Sources

Sources: Perplexity blog; AI Insiders, November 5, 2026. Note: the accuracy figures in this article are from Perplexity's own tests, with no independent run reported.

Related reading on this site: What Gets Cited by ChatGPT: Ahrefs' 1.4 Million-Prompt Study, Wikipedia Leads ChatGPT Citations at 10.1% in Ahrefs' November Ranking, Three AI Models Shopped Shopify's Catalog and Each Picked a Different Winner, Google CEO Pichai Sees Search Absorbed Into an Agentic Layer, 88% Disconnect Between Google Top 10 and ChatGPT Recommendations, Study Finds, and Google, Amazon, Microsoft and Meta Could Dominate AI Agents, Open Markets Warns.

More from Linkbaits

Turn this playbook into links

Linkbaits helps you create content that earns citations and backlinks. Start with the toolkit and build your first linkable asset this week.

Start free trial