Perplexity

Someone pushed back on the claim that Perplexity always searches, so I went and tested the obvious hole in it. If answers were cached, the search step would be theatre.

They aren’t. I ran two queries twice each, minutes apart, same logged-in Pro account, and caught all four on the wire.

Every run fired SEARCH_WEB with skip_search false. Nothing was served from a cache, and the answer prose was rewritten each time rather than replayed. On “what is technical seo” the two answers came back at 2,132 and 2,126 characters, opening “crawl, understand, and index” the first time and “crawl, understand, index, and rank” the second.

The interesting part is what happened underneath.

Head term, same set, different order. Both runs on “what is technical seo” pulled the same 10 domains. The order moved though, with Backlinko going from sixth to third between runs a couple of minutes apart.

Long tail, different set entirely. “best kettlebell weight for a third floor flat with thin floors and noisy neighbours” returned 30 sources on the first run and 15 on the second, with only 14 in common. The first run pulled 5 YouTube videos, an Instagram post and a handful of retailers that the second run never saw.

So the retrieval isn’t cached and it isn’t deterministic either. On a head term the candidate set is stable and only the ordering wobbles. Out on the long tail, half of it can turn over between two runs a minute apart.

Worth knowing before you trust any single Perplexity check, your own or a tool’s.

One schema note for anyone replicating this. Sources have moved since the July capture. They now sit at blocks[].web_result_block.web_results rather than at the top level, so an older extraction path returns nothing.

I’d already written the negative into the article. No per-source quality signal reaches the browser, and source ranking happens server-side where you can’t see it. June had no trust signal anywhere in the stream. An older free-tier capture carried a trust field on every source sitting empty, then build 7fe6ad4 dropped it completely.

Re-ran everything on 21 July, build df49f17, and it’s back with values in it.

"trust": {
  "level": 1,
  "name": "credible",
  "description": "is credible for first-party information about Discount Tire's U.S. tire and wheel retail stores, services, warranties, and related offerings."
}

Two tiers showed up, level 1 credible and level 2 trusted, the second sitting on goodyear.eu for “official Goodyear tyre product information”.

The description is the interesting part, because it’s a sentence rather than a score. caranddriver.com is credible for “long-established, professionally edited” automotive coverage. aaa.com is credible for “official information about AAA’s own membership services”. A domain gets believed about its own patch, never in general.

Coverage looks like a hand-curated registry working down from the head of the web, not a number computed per URL on demand. On the how-to run, 6 of 15 sources carried an entry and every one was a big official domain. The 6 YouTube results had nothing. Neither did any of the 15 sources on a local run made up of small editorial sites.

Two builds in two months took this field from empty to gone to populated, so read the tier names and the wording as a system mid-rollout rather than an API you can plan against.

Two changes on build df49f17, and a June capture script survives neither.

The answer socket doesn’t close any more. It goes quiet when the answer finishes and stays open, so anything waiting for the stream to end waits forever. That’s what quietly killed the first version of my script. You have to read the stream as it arrives instead, which is how the one in the teardown now works.

Then the fan-out. Re-running the commercial query, that single SEARCH_WEB step carried three queries where June had one. The verbatim string plus two close variants. One of them came back as “AI SEO tools Dubai pricing”. My city, folded into the expansion without being asked.

It’s still nothing like ChatGPT firing forty sub-queries and inventing brand names I never mentioned. The head query is your literal phrasing. But the gap is closing.

Local moved too. Same 0.996 on the places head, except the map engine got promoted and ran first this time, three rewritten variations ahead of the web search instead of after it. That session hadn’t shared location, the Places tab came back with “No places match this query”, and with no place-entities to cite the citations dropped straight back to the web results, the same class of listicle that got blanked in June.

Between them the two runs bracket the whole mechanism. When the places index delivers, businesses take every citation and the roundups watch. When it can’t, the roundups inherit the answer.

Perplexity works out what kind of question you asked before it searches. Then it sends you the decision, and the working-out with it.

Sixteen classifier heads ride on every answer. Places, video, shopping, image generation, a finance card, one head each, and all of them get a probability and a fixed bar to beat.

Ask it to explain the TLS handshake and the query files under TECHNOLOGY/CYBERSECURITY at 0.73 confidence. Nothing fires. image_preview climbs to 0.318 against a 0.42 bar, so an image strip nearly turned up and then didn’t.

The thresholds are the part worth writing down. image_generation wants 0.98, places_search_intent 0.85, shopping_intent 0.80, video_preview 0.50, calculator_widget 0.30. They were identical across all 7 queries in June and identical again on a different build 26 days later, so that’s a real decision boundary rather than a per-query mood.

ChatGPT gives you one label per query and stops there. Perplexity shows its working on every surface it considered, on every single query.

Method and the capture script are in the teardown.

The best thing in the ChatGPT teardown was the text bucket. Certain queries, mostly how-tos and definitions, get answered straight from training with an empty network tab. You can write the best page in the world and it’ll never be fetched.

Perplexity doesn’t have that gear. skip_search came back false on all 7 queries, including “how do I change a flat tyre step by step”, which is the exact question ChatGPT answered from memory without touching the web.

What it does instead is change mode without telling you. The flat-tyre query didn’t run as a normal search at all. It escalated to Study mode, model: pplx_study, search_mode: STUDY, and fired a video answer tab on top.

Same instinct as ChatGPT, opposite instinct about you. ChatGPT decided it already knew and shut the door. Perplexity decided to teach you and opened a video.

Re-ran it on the July build and got the same escalation, same video tab, video head at 0.988.

ChatGPT cites Reddit constantly and almost never cites YouTube. The reason is dull. It fetches a YouTube page, gets metadata instead of a transcript, and has no text to bind a citation to. Reddit is nothing but text, so Reddit gets quoted.

Perplexity runs the other way entirely.

On “best noise-cancelling earbuds under $150” it pulled 10 sources and cited 2 of them, YouTube and a niche eartips brand’s review page, 38 citations each. Three Reddit threads sat in the retrieved set and got cited zero times. The flat-tyre how-to cited YouTube 22 times across 3 videos.

Treat the counts as directional. One account, one city, a handful of queries I picked myself. The direction has a mechanism under it, so trust the shape and throw the numbers away.