Perplexity Cites the Same Pages Every Day: A 14-Day Zero-Variance Replication

What We Tested

E027 ran the E014 universal query set — ten queries, five Tier 1 proprietary concepts and five Tier 2 category terms — through Perplexity sonar-pro every day from 24 April to 7 May 2026. One run per query per day, 10:00 UTC, full annotations[] array captured per response. No content changes to cited pages during the window. No publishing freeze — all other site activity was logged.

The experiment extends E016’s 5-day noise floor measurement across a 3× longer window spanning two full weekly cycles. E016 found Perplexity citation rate at 40% ±0pp across five days. The question E027 asks is whether that zero variance holds when the window is long enough to capture weekday-of-measurement effects and index refresh cycles.

It does. But the more precise finding isn’t about the count — it’s about which pages get cited.

14-Day Perplexity Results

† Dual URL: old slug /retrieval-probability-geo-stack/ appeared alongside /retrieval-probability/ in the annotation array. 301 deployed 30 Apr; old slug persisted 8/14 days (Perplexity index propagation lag — see failure modes). ‡ Q09 System Memory: cited /geo-stack/ — a thegeolab.net citation but the wrong page for the query. Other days: GeoSCADA industrial software or Dell RAM hardware as annotation [1]. § Q03 LLM readability: first cited Day 7 after PerplexityBot re-crawled the page. Appeared 5/8 subsequent days.

The Core Finding: Zero-Variance Citation Identity

E016 reported aggregate stability: Perplexity citation rate = 40% ±0pp over five days. That framing is correct but it underspecifies what’s happening. E027 reveals the more precise finding.

Three query→page bindings were stable across every one of the 14 days:

• Q02 (Retrieval Probability) → /retrieval-probability/ — 14/14 days
• Q05 (GEO Stack) → /geo-stack/ — 14/14 days
• Q06 (Extractability) → /extractability/ — 14/14 days

Zero flip rate. No weekday-of-measurement pattern. The same pages, for the same queries, every day across two full weekly cycles. This is citation-identity stability — a stronger instrument property than citation-rate stability.

Why does citation-identity stability matter more than citation-rate stability? Because it changes what counts as a detectable result. If citation rate could drift ±20pp while citing the same pages, you’d need large sample sizes to detect treatment effects. Instead, any change in which page is cited is immediately interpretable — the instrument is precise enough to register a single-page movement.

The 22.0pp interpretability threshold established by E016 remains valid and is confirmed by this result. Zero variance persists over 14 days, ruling out the artefact of a short window.

Two Failure Modes of the Deterministic Zone

The deterministic zone isn’t universal. Two queries fell outside it, for different reasons. Both are characterisable — which means they’re avoidable in future experiment design.

Failure mode 1: Disambiguation miss (Q09)

Q09 asks about System Memory in GEO. Across 14 days, it produced four distinct outcomes as annotation [1]:

Dell RAM is the dominant annotation [1] at 6/14 days. The query “System Memory in GEO” is contested by four unrelated referents: computer hardware (Dell), industrial software (Schneider GeoSCADA), geospatial mapping (GeoServer), and the GEO Stack framework layer. None stabilises.

This is the expected result for a query where the key term (“system memory”) is a general-purpose technical phrase with established referents in multiple domains. The GEO Lab’s proprietary use has no Wikidata disambiguation to distinguish it from the established meanings. Until entity disambiguation is resolved — a target for E025 Wave 2 — Q09 cannot be used as a reliable measurement query.

Failure mode 2: Redirect propagation lag (Q02)

The 301 redirect from /retrieval-probability-geo-stack/ to /retrieval-probability/ was deployed on 30 April. Despite this, the old slug appeared in Q02 annotation arrays on 8 of the remaining 10 days. The propagation was non-monotonic — the old slug disappeared on Days 7, 8, and 13, then reappeared.

The redirect was never broken server-side (confirmed: curl -I returns 301 on both slash variants throughout). The persistence was entirely Perplexity index staleness. Minimum propagation latency on this domain: 7+ days, non-monotonic.

Practical rule: deploy any URL change at least 14 days before opening a measurement window, and monitor annotation arrays for old-slug persistence before declaring the instrument clean.

The Q03 Entry Event and What It Reveals

/llm-readability/ was not cited on any of Days 1–6. On Day 7 (30 April), PerplexityBot crawled the page at 10:00:26 UTC — confirmed from Nginx logs. The page was cited for the first time in that same morning’s cron run. From Day 7, it appeared on 5 of the remaining 8 days.

At the time, this looked like it might be an early signal from E003 — a heading-format intervention deployed on 18 April, 12 days before the crawl. But E003 R1 (run 2 May) returned a null result: /llm-readability/ was not cited on any platform, and Nginx logs confirmed PerplexityBot had not re-crawled the page between Day 7 and the R1 measurement.

The conclusion is clean: the Day 7 citation was a crawl-timing artefact. The re-crawl on 30 April refreshed Perplexity’s index entry for the page, producing a 5/8-day citation window that then expired. The heading-format intervention had no effect — not because it was inadequate, but because it was never observed by PerplexityBot during the measurement window.

This finding changes the E003 protocol for any future replication. The intervention was deployed. The crawl didn’t happen in time. The null result is structural, not content-driven.

T2 Null: No Authority Gate Signal at Day 14

T2 category queries returned zero citations across all 70 checks (14 days × 5 queries). No E025 Wave 1 entity-graph signal detected within the 14-day window.

This is the expected result. EDX established that below the domain authority gate, on-page variables and entity-graph interventions are non-operative on competitive category queries. E025’s pre-registered measurement horizon is M1 (22 May) and M2 (19 June) — not a 14-day window. The E027 T2 null doesn’t falsify E025; it’s a baseline reading within a window too short to detect entity-graph propagation effects.

Zero T2 citations across 70 checks also confirms the T1/T2 tier model holds over 14 days. Citation or nothing — no “mentioned” state was observed on any T2 query at any point in the window.

Four Rules for GEO Experiment Design

E027 produces four practical rules for anyone running controlled experiments on Perplexity:

1. Use Perplexity sonar-pro as the primary signal platform for T1 proprietary-concept experiments. Citation-identity is deterministic for disambiguation-resolved proprietary queries above the authority gate. Any movement in the citation set is signal, not noise.
2. Confirm disambiguation before the measurement window opens. Run target queries manually and verify annotation [1] reflects the intended concept. If a non-target-domain source leads the annotation array on multiple manual runs, the query is outside the deterministic zone until entity disambiguation is resolved.
3. Confirm PerplexityBot crawl before declaring an intervention active. Grep Nginx access logs for PerplexityBot hits on every target URL after intervention deployment. Do not open the measurement window until crawl is confirmed. This is a gate, not a check.
4. Deploy URL changes at least 14 days before any measurement window. Redirect propagation in Perplexity’s index is non-monotonic and incomplete within 14 days on this domain. Monitor annotation arrays for old-slug persistence before the window opens.

Frequently Asked Questions