AI citation rate: manual cell-level owned-link formula
This guide's manual cell-level owned-link rate is the number of eligible completed prompt-surface-repeat cells with at least one qualifying UI-marked citation or supporting-source link to a registered owned domain, divided by all eligible completed cells. It is not RankEcho's product Citation rate: the current product scorecard rolls up by prompt and can count a target URL or a detected brand or domain mention, while this worksheet requires a qualifying owned source link in each registered cell. The current Proof Loop also does not enforce this guide's common-cell intersection, so neither product output reproduces this measure. Use the manual audit worksheet with companion protocol and source-role ledgers, then report through the proof report. Keep unavailable and failed cells outside the denominator, compare periods on the common observed-cell intersection, and show full-window coverage separately. Site traffic is a separate outcome and is not required to calculate this measure.
What is the scope of this manual cell-level owned-link rate?
Use this metric for a declared prompt panel, services and surfaces, locale, account state, matching rule, and collection window. It summarizes observed cells under that contract; it is not a population probability, a universal engine benchmark, a rank, an impression metric, a referral rate, or a causal effect.
Provider dashboards can use different units. Bing reports provider-defined displayed-source citations and sampled grounding queries; Google Search Console reports link impressions from named Google Search features. Keep those aggregates separate from this guide's manual prompt-and-surface cells.
How does this differ from the RankEcho product metric?
RankEcho's current headline product metric uses its own parser and a prompt-level rollup: a target URL or a detected brand or domain mention can satisfy the product citation state for a prompt. This manual measure excludes mention-only cells and counts each eligible prompt-surface-repeat cell separately.
RankEcho's product proof paths use different records. The tracked-gap path starts from the gap audit's stored zero baseline for the prompt and summarizes engine-by-run observations collected at or after shipment. A commissioned-prompt Proof Card can split its own engine-by-run observations before and after ship time. Neither path stores these manual companion protocol and source-role ledgers or requires their exact common-cell intersection. Recompute the manual measure from the audit rows and companion ledgers, then treat this page as a separate manual protocol—not a description of a product report or import format.
What is an eligible observed prompt-and-surface cell?
A cell is eligible when its prompt, service, surface, locale, account or session state, model label when exposed, time window, and source-inspection procedure match the predeclared protocol. It is observed only when an eligible answer and its visible source locations were captured successfully.
Freeze a link-role rule for each provider and surface before counting. Include only links that the interface identifies as citations or supporting sources; retain merely related, relevant, or suggested links separately. Then resolve redirects, apply the declared host and subdomain rule, and retain the UI label plus original and resolved URLs. If the interface does not distinguish supporting links from related links, call the result a target-owned source-link rate rather than a citation rate. One cell contributes at most one numerator event even when it shows several qualifying target-owned URLs.
How do availability states affect the denominator?
Publish the state counts beside the rate. Never silently turn uncollected or unobservable cells into observed negatives.
| State | Observed denominator? | Required treatment |
|---|---|---|
| Planned | No | Report as not yet collected |
| Unavailable | No | Report the service or surface reason |
| Failed | No | Report the collection or parsing error |
| Observed without target-owned citation | Yes | Count once in the denominator and zero in the numerator |
| Observed with one or more target-owned citations | Yes | Count once in both numerator and denominator |
What is the manual cell-level owned-link formula?
Manual cell-level owned-link rate = eligible observed cells with at least one qualifying UI-marked citation or supporting-source URL that resolves to the registered target host / all eligible observed cells. Show it as n/N and a percentage. Declare link-role eligibility, URL normalization, rounding, and whether repeated answers are separate registered cells.
If the question is whether any source appeared, calculate a separately named any-source citation-bearing-answer rate. If the question is whether the target was mentioned or recommended, calculate separately named mention or recommendation measures. Never blend these outcomes into the citation numerator.
If the eligible observed denominator is zero, report the rate as not available and show the state counts. The expression 0/0 is undefined and must not be displayed as 0%.
What does the worked synthetic rate artifact show?
SYNTHETIC EXAMPLE — fictional brand, surfaces, sources, and observations; not RankEcho data, customer data, or a provider result.
| Component | Synthetic count | Calculation or treatment |
|---|---|---|
| Planned cells | 12 | Declared panel size |
| Unavailable cells | 1 | Disclose; exclude from observed denominator |
| Failed cells | 1 | Disclose; exclude from observed denominator |
| Eligible observed cells | 10 | 12 planned - 1 unavailable - 1 failed |
| Observed cells with an owned citation | 2 | Binary numerator under the frozen host rule |
| Baseline cell-level owned-link rate | 2/10 = 20.0% | Report numerator, denominator, and rounding together |
| Later full-window rate | 4/11 = 36.4% | One formerly failed cell is now observed and cited |
| Primary common-cell comparison | 2/10 to 3/10 = +10.0 percentage points | Same ten observed cell IDs in both windows |
Which synthetic cell ledger produces the rates?
The same fictional Northstar S1 ledger used by the gap analysis exposes all 12 cell IDs. Baseline has ten common observed cells, one unavailable cell, and one failed cell. In the later window, N05 changes to an owned citation and formerly failed N08 becomes observed with an owned citation; N08 is excluded from the common-cell comparison.
The owned flag is binary per cell. Other-source, competitor-replacement, related-link, mention-only, and no-source rows remain observed negatives for the manual cell-level owned-link rate; unavailable and failed rows do not enter that window’s observed denominator.
| Cell and context | Baseline state and link role | Later state and link role | Common observed cell? |
|---|---|---|---|
| N01 · P1 category comparison · Surface A/R1 | Observed · qualifying target-owned citation or supporting source | Observed · qualifying target-owned citation or supporting source | Yes |
| N02 · P1 category comparison · Surface A/R2 | Observed · other source | Observed · other source | Yes |
| N03 · P1 category comparison · Surface B/R1 | Observed · no visible source | Observed · no visible source | Yes |
| N04 · P1 category comparison · Surface B/R2 | Unavailable | Unavailable | No |
| N05 · P2 calibration interval · Surface A/R1 | Observed · other source | Observed · qualifying target-owned citation or supporting source | Yes |
| N06 · P2 calibration interval · Surface A/R2 | Observed · no visible source | Observed · target-owned related link · excluded from citation numerator | Yes |
| N07 · P2 calibration interval · Surface B/R1 | Observed · no visible source | Observed · no visible source | Yes |
| N08 · P2 calibration interval · Surface B/R2 | Failed | Observed · qualifying target-owned citation or supporting source | No |
| N09 · P3 integration requirements · Surface A/R1 | Observed · qualifying target-owned citation or supporting source | Observed · qualifying target-owned citation or supporting source | Yes |
| N10 · P3 integration requirements · Surface A/R2 | Observed · mention only | Observed · mention only | Yes |
| N11 · P3 integration requirements · Surface B/R1 | Observed · other source · competitor replacement | Observed · other source · competitor replacement | Yes |
| N12 · P3 integration requirements · Surface B/R2 | Observed · no visible source | Observed · no visible source | Yes |
How do the synthetic surface rates reconcile?
Calculate the full-window and common-cell summaries from the same rows. Surface totals must add to the all-surface numerator and denominator in each column.
| Surface | Full window: baseline → later | Common cells: baseline → later | Matched change |
|---|---|---|---|
| Surface A | 2/6 = 33.3% → 3/6 = 50.0% | 2/6 = 33.3% → 3/6 = 50.0% | +16.7 percentage points |
| Surface B | 0/4 = 0.0% → 1/5 = 20.0% | 0/4 = 0.0% → 0/4 = 0.0% | +0.0 percentage points |
| All surfaces | 2/10 = 20.0% → 4/11 = 36.4% | 2/10 = 20.0% → 3/10 = 30.0% | +10.0 percentage points |
How do you calculate matched change across periods?
Form the intersection of cells observed in both periods under the same prompt, service, surface, locale, account state, matching rule, and other controllable protocol fields. Calculate each period's rate only on that paired set, then subtract the earlier percentage from the later percentage.
In the synthetic common set of ten observed cells, the earlier period has 2/10 owned citations and the later period has 3/10. The descriptive matched change is +10.0 percentage points, from 20.0% to 30.0%. The later full-window rate is 4/11 = 36.4% because one formerly failed cell becomes observed and cited; that changing availability is why the full-window difference is not the primary like-for-like comparison.
How should matched and available-case summaries differ?
The matched summary answers what changed among cells observed in both periods. An available-case summary uses every observed cell in each period and can change because the available mix changed. Publish it separately, with its own counts, when it is useful.
Do not fill missing cells with zero unless the protocol explicitly defines and labels a planned-panel measure. Do not compare one provider's citation count with another provider's link impressions as though they share a numerator and denominator.
Which deterministic validation checks must pass?
The calculation is complete when another person can reproduce every count from the retained cells. It does not wait for a traffic threshold, ranking change, referral, or future citation.
| Required check | Pass condition | Evidence to retain |
|---|---|---|
| Protocol | Prompt panel, services, surfaces, locale, account state, window, link-role rule, and host rule are frozen | Versioned protocol and cell registry |
| Availability | Planned, unavailable, failed, and observed counts reconcile | State table with reasons |
| Numerator | Every positive has a qualifying UI-marked citation or supporting-source URL resolving to the target-owned host | Raw answer, UI label and role, source location, original URL, resolved URL, and host match |
| Arithmetic | Displayed n/N, percentage, rounding, and state totals recompute exactly | Machine result and independent recomputation |
| Matched comparison | Both periods use the same observed-cell intersection and expose gains, losses, and unchanged cells | Paired cell table and set-difference checks |
| Interpretation | The result is labeled descriptive and keeps alternative explanations visible | Published limits and semantic check |
Which interpretation risks should remain visible?
A larger rate can reflect answer variability, surface availability, source-display changes, prompt composition, account or locale differences, model changes, or measurement error. A small fixed panel is useful for its declared cells but does not justify a confidence band or population claim without an appropriate sampling design.
A citation is not automatically a recommendation, endorsement, accurate claim-level source, click, referral, conversion, or rank. Open the source, preserve the answer context, and report those outcomes separately when observed.
How does the manual cell-level owned-link rate close the Find-Fix-Prove loop?
Link each matched result to the original gap classification and the exact deployed change. Report the deterministic content acceptance result beside, but not merged with, the later answer observation.
No movement and adverse movement remain valid results. Use them to revise the next diagnosis; do not keep editing until a favorable rate appears or describe timing alone as causation.
Who maintains this manual owned-link-rate guide?
Written and maintained by Abiot Y. Derbie. Published 2026-09-07; sources checked 2026-09-06; last updated 2026-09-07. The rate and matched-change examples are synthetic and do not represent customer outcomes.
Recheck provider source displays, dashboard definitions, host-matching rules, and arithmetic before applying the metric. Send material corrections through the contact page.
Sources reviewed
Provider eligibility and measurement claims below were checked against primary documentation. These records do not establish a universal selection formula, causation, or a guaranteed ranking, impression, recommendation, or citation.
4 claim-level source records
| Claim reviewed | Official source | Review record |
|---|---|---|
| Bing's AI Performance report counts citations displayed as sources, reports cited-page activity, and exposes sampled grounding queries. Microsoft says those counts do not indicate placement, ranking, authority, page importance, or a page's role in an individual answer. | Bing Webmaster Blog: AI Performance | Checked 2026-09-06 · Public preview announced February 10, 2026 · Primary-source documentation check; Bing's provider-defined aggregate is kept separate from this guide's registered prompt-and-surface cells and from other providers' reporting units. · Confidence: High |
| Google's Generative AI performance report counts link impressions from AI Overviews and AI Mode and groups data by page, country, date, or device. Its documentation does not expose prompts, citation causes, a universal rank, or a source-selection formula. | Google Search Console: Generative AI performance report | Checked 2026-09-06 · Worldwide rollout stated as August 31, 2026 · Primary-source documentation check; a Google Search link impression is not relabeled as a fixed-cell citation, mention, recommendation, or matched prompt observation. · Confidence: High |
| Google says Gemini Apps may show inline links and a Sources button, but not every response has source links and a listed page can relate to only part of a response. | View related sources from Gemini Apps | Checked 2026-09-06 · Current Gemini Apps source guidance · Primary-source documentation check; a visible Gemini source is recorded only for the observed response and is not treated as a durable rank or claim-level support for the whole answer. · Confidence: High |
| OpenAI says ChatGPT responses that use web search may include citations and may expose cited sources and other relevant links in a Sources view. It also warns that search results and citations can be incomplete, outdated, or incorrect and should be checked against the opened source. | OpenAI Help: searching the web with ChatGPT | Checked 2026-09-06 · Current ChatGPT search source guidance · Primary-source documentation check; the observed visible citation is retained as evidence for that response, while source support and answer accuracy are checked separately. · Confidence: High |
Frequently asked questions
All eligible observed prompt-and-surface cells under the declared protocol. Planned, unavailable, and failed cells are disclosed separately and are not silently counted as observed negatives.
No, not for the cell-level rate defined here. The cell contributes one positive numerator event. A separately named URL-count metric can preserve the three links without changing the cell rate.
No. Some interfaces can mix supporting citations with related or relevant links. Freeze the eligible UI roles per surface, retain the rest separately, and use the broader name target-owned source-link rate when the interface does not distinguish them.
No. The product metric rolls up by prompt under its parser and can count a target URL or a detected brand or domain mention. This manual rate requires a qualifying owned source link per eligible cell and excludes mention-only cells.
Use the intersection of cells observed in both periods under the same protocol, show each raw n/N, and report the percentage-point difference. Publish available-case results separately if needed.
No. It is a descriptive change under the matched protocol. Timing supports an association, while answer variability and other changes remain plausible unless the design isolates them.
No. The metric uses eligible observed answer cells and preserved visible sources. Traffic is a separate later outcome and does not determine whether the arithmetic is reportable.