Perplexity source selection works through at least two visible layers: a 16-head intent classifier (an automatic routing system that runs before any search begins) and an emerging trust registry that attaches scoped credibility notes to specific domains. How Perplexity ranks sources within those constraints remains server-side and invisible to outside observers.
That picture comes from a reverse-engineering report by Search Engine Journal. The researcher intercepted Perplexity’s live data stream on a single logged-in Pro account in Dubai, capturing eight query sessions on June 25, 2026 (build 7fe6ad4) and re-running three spot checks on July 21, 2026 (build df49f17). Structural findings, such as which fields exist and how they behave, are high-confidence. Percentage figures and frequency counts are directional only given the small sample.
What does Perplexity’s 16-head classifier actually decide?
Before running a single search, Perplexity scores the query across 16 possible intent surfaces, called heads, each with a fixed probability threshold. The entire scorecard ships to the user’s browser in a field called classifier_results.mhe_predictions_full, which is more routing logic than most AI platforms expose publicly.
The 16 heads cover distinct surface types: local places, shopping, video, weather, image generation, finance cards, and others. Each head returns a probability, a threshold, and a true/false firing decision. Atop the heads sits a domain_subdomain label, Perplexity’s topic taxonomy for the query. Per the SEJ report, thresholds were identical across all seven query types in June and unchanged 26 days later on a different build. Domain labels shifted with the query as expected:
- “best AI SEO tools 2026” classified as TECHNOLOGY/ARTIFICIAL_INTELLIGENCE at 0.86 confidence
- “Ahrefs vs Semrush” classified as BUSINESS/DIGITAL_MARKETING at 0.94 confidence, the highest score in the capture run
- “latest Google algorithm update” scored TECHNOLOGY/INTERNET_TECHNOLOGIES at 0.49, the lowest, because news queries resist a single clean topic
The classifier output determines which surface a brand is actually competing for. A “best X near me” query clears the places threshold and opens a maps contest, not a standard citation fight. A how-to query pulls a video tab alongside text results.
Does Perplexity always search the web, or does it sometimes answer from memory?
Perplexity searches the web on every query. The skip_search flag (a signal that tells the system to answer from training data without making any web request) was false on all seven query types in the capture, including step-by-step instructional queries that ChatGPT, per the same researcher’s prior teardown, resolves entirely from memory.
This structural difference matters for content visibility. In ChatGPT, a how-to query can resolve from training with no external page cited at any stage. In Perplexity, every how-to query fetches from the web. The SEJ capture found the flat-tyre query escalating to Study mode (model: pplx_study), a teaching-oriented mode, and triggering a video tab with the video classifier head scoring 0.988 on re-run in July. That behavior was identical across two builds roughly four weeks apart.
What is Perplexity’s trust field, and what does it actually measure?
As of the July 21, 2026 capture, Perplexity attaches a trust object to some domains in its results, carrying a numeric level, a tier name, and a written scope sentence describing what the domain is trusted for. The field was absent in the June capture and absent-but-present-empty in an earlier free-tier capture, meaning it has changed state at least three times across two months of builds.
The SEJ report found two tiers:
- Level 1, “credible”: for example, caranddriver.com, described as credible for “long-established, professionally edited” automotive coverage
- Level 2, “trusted”: for example, goodyear.eu, described as “trusted for official Goodyear tyre product information” including its EMEA and fleet business
Every captured trust entry followed a first-party pattern: a domain is trusted about its own products, services, and operational area, not trusted in general. Scope sentences describe what a domain owns, not how large or well-known it is.
Coverage was uneven in the capture. On the how-to query, 6 of 15 sources carried a trust entry, all large official domains. YouTube carried no trust entry on any of its citations on one query, yet still accounted for 14 of that query’s 40 cited sources. A missing trust entry did not prevent citation. The researcher explicitly flags the tier names and wording as “a snapshot of a system mid-rollout, not a stable API.”
How deep does Perplexity search before generating an answer?
Perplexity’s default search fan-out (the number of search steps it runs before composing an answer) is shallow. For six of the seven queries in the capture, it executed a single SEARCH_WEB step, sending the user’s query near-verbatim to the web and retrieving a compact result set. “Ahrefs vs Semrush” went out unchanged; “best AI SEO tools 2026” went out unchanged and returned 10 results.
The SEJ researcher contrasts this with ChatGPT’s behavior on the same queries. Perplexity’s default behavior is one query, one result set, one answer.
What this means for AI-search visibility
The clearest way to read these findings: Perplexity is building a trust registry that works like a specialist directory, not like a general credibility score. Think of it less like a domain authority number and more like a verified badge that says “this site is the official source on X.” The optimization question is not “how do I appear authoritative overall?” but “what specific topic does my domain unambiguously own?”
The goodyear.eu example illustrates this concretely. That domain earned a Level 2 “trusted” designation for Goodyear tyre products specifically, not for tyres in general. A mid-size brand with deep, well-structured documentation on its own product line is a stronger first-party candidate for a trust entry than a large general publisher with broad topical coverage. Size appears to matter less than specificity of ownership.
The YouTube finding adds a second, equally important layer. YouTube collected 14 of 40 cited sources on one query with zero trust entries attached. This suggests Perplexity is operating at least two parallel systems: a trust registry that is curated, first-party, and still rolling out, and a separate citation-volume system driven by relevance and content-format signals the stream does not expose. A missing trust entry is not an exclusion. It means a domain is competing on the second track only, without the credentialing signal of the first.
From a Hingewise assessment perspective, the classifier scorecard is underreported relative to the trust story in most GEO (Generative Engine Optimization, meaning optimization for AI-generated answers) discussion. The trust field may look different next month; it has already changed state three times. The classifier thresholds, by contrast, were stable across two builds 26 days apart. That stability makes the classifier output more actionable right now: it tells you which competitive surface your priority queries land in, maps, video, shopping, or standard citations, before you write a single piece of content. Most brands have never looked at it. That is the gap worth closing first.
Before assuming your site is visible in Perplexity, check these
- Run your priority queries in Perplexity and note which surface fires: a map card, a video tab, or standard cited results. The classifier output tells you which competitive slot you are actually in.
- Check whether your domain appears in Perplexity answers at all, across both standard and Study mode queries.
- Identify what your domain is the unambiguous first-party source for: your own products, your own data, your own services. That is the scope the trust field appears to describe.
- If how-to or instructional queries are a priority, note that Perplexity always fetches for these. There is a page slot and a video slot to compete for, unlike in ChatGPT where the same query may resolve from memory.
- For news queries, expect lower classifier confidence and a more contested citation pool. The SEJ capture scored the news query at 0.49, the lowest of the seven query types.
- Do not treat a missing trust entry as an exclusion signal. YouTube had zero trust entries and dominated citations on one query in the capture.
The trust field’s rapid state changes (present, absent, present-with-values across two months) are the key variable to watch in the months ahead. Whether Perplexity expands the registry to smaller editorial domains, whether the tier structure changes, and whether scope sentences begin reflecting topical authority beyond first-party ownership remain open questions. The classifier routing logic appears settled. The trust credentialing system, by contrast, is still being built.
Source: Search Engine Journal, “How Perplexity Actually Picks Sources (I Read The Stream, Not The Answers).” Capture data from builds 7fe6ad4 (June 25, 2026) and df49f17 (July 21, 2026). Single user, logged-in Pro account, Dubai geo, 8 captures across 7 query types.
Lam Nguyen · Hingewise