The Algoholic Index · Q3 2026 · original-data report

How visible is anyone, really, inside the answer?

The Algoholic Index is the quarterly measure of generative-engine visibility — who gets cited, how faithfully, and how durably, across verticals and the frontier-model panel. This is the methodology edition: it preregisters exactly what the Index measures and how, so the numbers, when they land, arrive against a fixed and public design.

Vol. I · Edition 01Published 2026-07-05CC BY 4.0Q3 dataset — pending probe run
Read this first — the honesty note. An original-data report is only worth the paper it is printed on if the data is real. The Q3 numbers require a controlled probing run against the frontier-model panel — live model access this project has flagged as an open dependency (model API keys + budget). Rather than fill the tables with plausible-looking figures, this edition publishes the method now and the measurements when the run completes, dated as a revision. A preregistered design that anyone can check — and later hold the numbers to — is the point. Inventing the numbers would be the exact practice this archive exists to stand against.

Why an index, and why quarterly

Classical search had a hundred visibility trackers because rank was observable — you could watch a position move. Generative retrieval has almost none, because the thing that now decides whether your finding reaches a buyer happens inside a model's forward pass, unranked and unlogged. The only way to see it is to probe it, repeatedly, on a fixed protocol, and report what comes back. That is what the Index does: a standing, quarter-over-quarter instrument for a layer the industry currently navigates by anecdote.

Quarterly, because the panel itself moves. Models ship, retrieval stacks change, grounding heuristics shift. A one-off snapshot is a rumor; a dated series is a trend line — and the trend line is where the citations, the press pickups, and the defensible authority live.

What the Index measures

Five dependent variables, each defined so it can be computed the same way every quarter. Precise definitions and the scoring rubric are the whole value of preregistering — they are what make Q3's number comparable to Q4's.

01

Citation share

Of the model answers that touch a claim, the share that attribute it to a given source. The headline visibility number.

02

Statement match

How faithfully a reproduced claim matches the source wording — verbatim, paraphrased, or absorbed uncited.

03

Citation persistence

Whether a citation survives re-probing over a 30-day window, or decays as the model's working context shifts.

04

Cross-model variance

How far the same claim's treatment spreads across the model panel — the instability a single-model check hides.

05

Answer concentration

How few sources the engines route a category's answers through — the winner-take-most curve of generative retrieval.

The method (preregistered)

Fixed in advance, so the run cannot be steered toward a preferred result:

  1. Corpus. A fixed set of buyer-stage query sets across 7 verticals, drawn from real decision-stage questions rather than head keywords. The query sets are frozen before the run and published with the dataset.
  2. Panel. The frontier-model panel spanning the three answer-engine classes — bulk-trained closed-book, search-grounded, and on-demand fetch — probed through their first-party APIs on identical prompts.
  3. Probe. Each query is put to each model on a controlled schedule; responses are captured verbatim and classified with the four-class citation rubric from the citation-behaviour taxonomy (verbatim cite / paraphrase-with-source / silent absorption / contradiction), plus a no-show class.
  4. Persistence. The same probes are re-run at days 7, 14, and 30 to measure citation decay — the metric a single snapshot cannot see.
  5. Scoring. Blinded classification with a documented rubric and inter-rater agreement reported, exactly as in the audit paper. Every figure traces to a saved response.

The verticals

The Q3 edition reports each metric per vertical, because citation behaviour is not uniform — regulated and YMYL categories concentrate answers on institutional sources far harder than horizontal SaaS does, and that difference is itself a finding.

  • SaaS / developer tools
  • Cybersecurity
  • Financial services
  • Healthcare / YMYL
  • E-commerce
  • Legal
  • Cannabis / regulated

What the Q3 tables will report

The dataset — when the probe run completes — populates a fixed set of tables and figures, released as raw data to /data under CC BY 4.0 so anyone can reproduce or contest them:

Preregistered structure — dataset pending the probe run
VerticalCitation shareVerbatim rateSilent-absorption rate30-day persistenceCross-model variance
SaaS / developer tools
Cybersecurity
Financial services
Healthcare / YMYL
+ 3 more verticals

Every cell above fills from a real, dated probe run — never estimated. The empty table is deliberate; it is a promise, not a placeholder for invented numbers.

Release cadence

The Index runs each quarter — Q3 2026, Q4 2026, Q1 2027 — building the citable series that turns a one-time report into an industry reference. Each edition: methodology fixed, run executed, dataset to /data, findings written up, raw responses available on request for replication. The methodology published here is the standing protocol; only the numbers change quarter to quarter.

Cite this

Sasson, G. (2026). The Algoholic Index — Q3 2026 (methodology edition). Algoholic. https://algoholic.com/reports/2026-q3

Want your category in the Index?

The Index measures categories in aggregate. The Citation Map and the LLM Visibility Audit run the same protocol on your specific domain and queries.

Run the audit →