How Citeon AI measures: 21 words, one definition each, from the code

The question and the answer

Question (prompt)

A buyer question asked to each AI engine, in its own language, from the workspace's country. Every question is asked every second scan (two halves on alternate days); a new or missed question joins the next scan.Computed in engine/rotation.py, lib/scan-schedule.ts

Answer

One engine's reply to one question in one scan. Stored in full with the sources it cites. An attempt that returned nothing is stored as an attempt with its reason, never as an answer.Computed in engine/store_snapshot.py, prompt_results, prompt_attempts

No answer

The latest attempt for that question got nothing back from the engine: a timeout, an error, or Google showing no AI Overview for that search. Never counted as absent, never as zero.Computed in prompt_attempts, lib/prompt-stats.ts

The four verdicts on an answer

Named (product)

A product's or collection's full title appears in an answer, whole words only; a one word title needs eight letters or more and navigation labels never count. Written when the answer is stored, with the sentence as evidence.Computed in engine/catalog.py hits_for_answer

Recommended

Named, and the judge reads the answer as presenting the brand as a pick, not a passing mention. For Google AI Overviews and AI Mode, where no judge runs: named and cited.Computed in engine/audit.py, engine/cutover_v2.py restate_row

Cited (product)

An answer links that product's or collection's page (locale prefix and query ignored). Cited pages are stored since 8 October 2026; earlier answers cannot show it.Computed in engine/catalog.py, prompt_results.urls

Absent

The engine answered, and the brand is neither named nor cited.Computed in lib/prompt-stats.ts

The rates and the score

Mention rate

Answers that name the brand, as a share of all answers in the window. The score tooltip calls it "how often AI engines name you".Computed in engine/audit.py (per scan), engine/rollup.py (rolling)

Recommendation rate

Answers that recommend the brand, as a share of all answers in the window.Computed in same

AI visibility score

0 to 100: 60% of the mention rate plus 40% of the recommendation rate, over the scans of the last 28 days, each day in the window weighted equally. Not the GEO score.Computed in engine/audit.py line "score = round(100 * (0.6 * mention_rate + 0.4 * rec_rate))", engine/rollup.py

Engine score

The same formula on one engine's answers only. Engines are independent scores, not shares of one total. An engine with no answer shows "not shown" (Google showed nothing) or "no answer", never 0.Computed in engine/rollup.py engine_scores

Against the competitors

Share of voice

Of every time an answer named one of the tracked brands (the brand and its tracked competitors), the share that named this brand. Example: the brand named 4 times, rivals 12 times in total, share 25%. It is a share of names, not a share of answers.Computed in engine/audit.py _share_of_voice

Rank (share of voice)

The brand's place among the tracked brands by share of voice, 1 = most named. Unknown when nobody was named in the scan; last when rivals were named and the brand was not.Computed in engine/audit.py _share_of_voice

Position

Among the tracked competitors named in the same answer, how many are named before the brand, plus one. 1 = named first. Only exists when at least one tracked rival is named in that answer.Computed in engine/matcher.py position

Rank among every brand named

The same count over every brand the judge listed in the answer, not only tracked rivals (Peec's documented rule). Shown on the row from ten ranked answers.Computed in lib/prompt-stats.ts rankAmongNamed

Citations

Citations

How many times the brand's own pages were cited across the scan's answers. Per engine on the Citations page.Computed in engine/audit.py client_citations

Time, corrections and storage

Window

28 days unless the page says otherwise. The Prompts page offers 7, 14 and 28.Computed in lib/prompt-stats.ts, app/dashboard/[section]/page.tsx

Restart

When the workspace's country is set or changed, the days asked from another country leave the average; the chart marks the day.Computed in engine/market.py comparable

Restatement

A stored answer re-read with the current method or names; the verdict can change, the answer text never does. Every restatement writes a line to the corrections ledger with the number of answers that changed.Computed in engine/cutover_v2.py, metric_corrections

Retention

Answers keep their verdict, their judge quote, their cited pages and domains for good. The answer text is kept for 120 days, then cleared; attempts are deleted after 120 days. An answer without text is never re-read.Computed in engine/store_snapshot.py, engine/cutover_v2.py

The site, not the answers

GEO score

The site audit's grade of the site's readiness for AI search (crawlability, structured data, content), on the Reports page. Not the visibility score: that one measures the answers.Computed in engine/geo_audit.py

Also defined

Catalogue

The shop's own collections and products, read from the shop each week (Shopify's public listing first). Collections a shopper can reach come first; merchandising and app collections are kept but never shown.Computed in engine/catalog.py, Prompts

Readiness (product)

Which of the fields the shopping surfaces need a product page's own markup gives: the OpenAI product feed's required fields and Google's merchant listing fields, required and recommended, with Google's title and description length rules. Read by the monthly site audit for the product pages in its sample; a product not in a sample yet says not checked. The shop's own name meets the feed's seller name.Computed in engine/product_readiness.py

Cross check (traffic)

GA4's AI Assistant channel shown beside Citeon's own tracking over GA4's own window, never instead of it. When the two differ by more than a third the page says why: GA4 can be blocked by browsers and extensions, app visits land in GA4's Direct, and GA4 key events are whatever the site defines where Citeon leads are contacts only.Computed in components/dashboard/TrafficPage.tsx

What AI sends people to

The pages AI visitors landed on, grouped by kind (products, collections, articles, home, other). A catalogue page shows its title, the stored answers that cite it and the tracked question with the most answers that touches it; a page outside the catalogue shows a dash for those two, not a zero.Computed in lib/landing-kinds.ts, Traffic

Authority

DataForSEO's backlinks rank on its 0 to 100 scale (based on the number of referring domains) and the referring domains count, for the site and up to ten rivals, measured monthly. One vendor's model of links; Google's representatives said in 2020, as reported by Search Engine Journal, that Google does not use domain authority. A domain the vendor has never crawled says so instead of showing 0. In Ahrefs' study of 75,000 brands in AI Overviews (26 May 2025, Spearman), branded web mentions correlated 0.664 with being shown and Domain Rating 0.326: a correlation, not a cause.Computed in engine/authority.py, AuthorityCard

Visited (product)

An AI session, by the tracking's session rule, landed on that page in the window. A dash when tracking is not installed; a real zero when it is and nobody landed there.Computed in lib/pulse.ts aiLanding

Asked (product)

A tracked question shares a word with the item's title (the same rule as the win plan), shown with the share of those questions' answers that name the brand. A dash when no question touches it.Computed in lib/catalog-signals.ts

Why the words are published

A score you cannot reproduce is a score you cannot trust. Every figure in the dashboard, in a free audit report and on the data page is one of these words applied to stored answers, and every answer keeps its verdict, its judge quote and its cited pages so the figure can be opened to the sentence it came from. When a definition changes, the stored answers are re-read and the change is logged with the number of answers it moved; the answer text is never edited.

One word, one definition, one place in the code. Not measured is empty, never zero.

Citeon AI, the method rule