GEO-095·Issue 08·Measurement & Growth·95 / 132
Measuring Off-Site Authority: Mention, Citation, Recommendation
Off-site mentions climb while AI's candidate list stalls. Three layers, separate denominators, a five-step log, and a quarter that bought reach, not a reason.
Issue focusOff-Site Authority: Digital PR, Communities, and the Knowledge Graph
In August, a US air-quality brand’s EU growth lead sits in a quarterly review. The agency deck lists a magazine feature, a trade roundup, two podcasts, and a mentions line rising for three quarters. That afternoon she asks an assistant, in German, which purifier suits a 70 m² Berlin Altbau flat with two cats and radiator heating. The answer names three brands, none of them hers. She asks which placement earned that answer. Nobody can say. The deck counts mentions; no row records an answer.
Off-site authority is three layers, not one: mentions are reach, citations are trust, recommendations are conversion. A single score hides the only layer that moves a shortlist.
Three actions, three kinds of evidence
A mention answers whether anyone knows you exist. A citation answers whether a model will use you. A recommendation answers whether a buyer should pick you. Different places, different recorders.
| Layer | What happens | Counting unit | Owner | Common miscount |
|---|---|---|---|---|
| Mention | Brand or model appears in off-site text | One appearance per independent source | PR, content | Syndicated copies counted per outlet |
| Citation | An answer credits a claim to your page, or to text speaking for you | One attribution per answer | Growth | The same answer rephrased, counted twice |
| Recommendation | You appear in the candidate set with a reason to compare | One ranking per answer | Content | Presence counted, comparison ignored |
Not an automatic funnel. A model can quote your spec sheet and still write “another brand also offers similar products”; it can name you as the counterexample. Mentioned is not retrieved, retrieved is not used, used is not recommended.
Add the layers into an authority score and the direction disappears. Mentions can be bought. A citation can be earned with one checkable fact. A recommendation needs a reason a buyer would repeat. Book growth in the first two against the third and the team believes the shortlist is improving while it stands still.
Different denominators, so do not add them
The mention denominator is the monitored source set. Add two forums, or a German market, and the count rises by arithmetic while authority does not move. The citation denominator is the frozen prompt set times samples; change the set and the rate stops being comparable the same day. The recommendation denominator is only those prompts that returned a candidate list; ask again next week and the list reshapes.
One variable gets skipped: aliases. Communities write model numbers, series names, abbreviations, misspellings. Searching the registered name alone misses the honest discussion, and a missed detection reads as “nobody mentions us.”
Log three layers separately, each with its own definition and trend. To combine, freeze the denominator first and compute a rate of change. Never sum three columns and call the result authority.
A five-step routine, denominator first
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Define the entity | Brand, model, series, abbreviations, misspellings in one alias table | Brand, content | Any alias returns the same sources |
| 2 Freeze the prompt set | 30–50 prompts by category, scenario, comparison, after-sales; language and market on every row | Growth | Unchanged three months; a change opens a new baseline |
| 3 Log mentions | URL, date, source type, stance; syndicated copies merged | PR, content | Ten sampled rows trace to the original |
| 4 Log citations | Re-sample; copy the answer and the source it credits | Growth | Every attribution points to a page |
| 5 Log recommendations | Candidate list, rank, stated reason | Content | You can answer “why did we miss” |
Step 5 is skipped because it is manual, yet it alone explains “citations up, orders flat”: you are absent, or fourth, with the reason “also offers.”
The monthly report answers three questions
Which layer leaks: mentions rising with citations flat means reach is not becoming liftable material; citations rising with recommendations flat means the material is not becoming a reason to compare. Which source type leaks: media, communities, reviews, directories each get a row, a type yielding mentions but never citations is a candidate to stop. What moves next month: which type loses budget, which gains it.
No total score appears: three trend lines, one table by source type, one action with an owner and a verification date.
Example: one quarter, three columns
Example — brand and figures are invented; only the accounting method is reusable.
A smart pet-feeder brand tracks 40 English prompts in the US, sampled every two weeks. Quarter start: 60 mentions, a 9% citation rate, 4 appearances in 40 candidate lists. Quarter end: 95 mentions, 10%, still 4.
Read apart: over half the 35 new mentions are copies of one press release — one source, not thirty-five; the one-point citation move sits inside sampling noise; the shortlist did not move, and the reason is always “also offers.”
The quarter bought reach, not comparison reasons. The next move is different material — test conditions, jam rate, disassembly steps — compressed into a table anyone can quote, the kind described under citable assets. Then move budget from wire distribution to reviews and named experts, ordered with the off-site investment ranking.
Where this breaks
Paid placements. Sponsored content is a mention, not trust; an undisclosed advertorial that gets recognized costs more than it buys. Its own column, never merged with earned mentions.
Syndication. Three outlets running one release are one source to a model. Deduplicate by source, not by link.
Citation is not endorsement. Quoted as the counterexample, or filed under drawbacks, still a citation. Keep a stance column: positive, neutral, negative.
Negative recommendation. “Not recommended for multi-cat households” deserves more attention than absence; it gets quoted long after the placement is forgotten.
Two cases do not fit. A brand with near-zero mentions should fix crawling and entity consistency first; a category closing almost entirely inside a marketplace should watch reviews and fulfillment instead. The ledger assumes someone off-site speaks for you — the division of labor is mapped in AI’s off-site trust network, and the funnel position is summarized on the measurement and growth hub.
This month, do one thing: take 30 real buyer questions, fix language and market, run them for two weeks, and log “mentions you,” “cites you,” “recommends you” as three columns you never merge. In week three, the ugliest column is the one to fix.