An answer names a company in a heading, repeats the trading name in its description and cites the company’s website. Another line uses the Arabic spelling. If you count every string, the brand appears four times. If the question is whether that brand was named in that answer, the answer is one. The citation is a separate observation, and the spelling variants need an identity check before they can be combined.
Use a consistent unit of counting and an evidence record for names that could refer to more than one business. AI Visibility’s monthly reporting guide explains what belongs in a recurring visibility report. This walkthrough focuses on how to produce defensible counts before those numbers go into the report. AI Visibility monthly reporting guide
Decide which kind of mention you are counting
Write the rule in terms a second reviewer can apply. A business might count as named when it has an entry in a list, when it appears anywhere in answer prose, or when the answer explicitly recommends it. Each rule captures a different kind of appearance. Keep a citation field even if the report also measures mentions. A source link can point to a directory, a publisher or a company that the text never names.
For a brand appearance count, use one row per answer and verified brand. Keep each original occurrence in the evidence log and deduplicate when calculating placements. If the same brand appears in three independent observed answers, it can count three times across the batch. The second and third occurrence within one answer remain useful evidence about context and prominence while contributing no additional placement.
| Observation | Record in the evidence log | Count under a one-brand-per-answer rule |
|---|---|---|
| Repeated spelling in the same answer | Every occurrence and its position | One verified placement |
| Approved alias for the same business | Original text and the matched identifier | Same placement as the main name |
| Distinct businesses in one entry | Each identifiable name | One per qualifying business |
| Domain citation with no qualifying name | Source URL and any verified owner | Citation channel only |
| Ambiguous name or branch | Original wording and reason for uncertainty | Pending unless resolved under the rule |
If the article says a company is recommended, define what qualifies for that label. A name can appear in an example or in a statement about who should be avoided. Counting those as mentions may be appropriate for a broad visibility measure. Calling them recommendations would change the meaning. Keep the placement type and its context in the source record so the report can be precise.
Build an alias map with evidence
Give each business a stable identifier in the coding sheet. Attach the official public name, approved abbreviations, Arabic variants and evidence that supports each association. For a local branch, keep the parent brand and branch identifiers in separate fields. Decide at which level the report will count. A chain-level result can consolidate several branch labels within one answer, while a branch-level result needs to preserve which location was actually named.
Use text normalization to prepare candidates for review. It can remove accidental differences in capitalization or spacing and reconcile equivalent Unicode representations. It cannot determine whether two businesses share a name or whether an Arabic rendering belongs to a particular company. Preserve the original wording, and treat the alias map as a separately maintained decision. Unicode’s normalization specification concerns text representation rather than business identity. Unicode normalization standard
Verify ambiguous candidates against authoritative information. The official website, branch page or confirmed brand profile can help identify the entity. A shared directory domain or a common word is weak evidence. Schema.org’s sameAs definition is a useful reminder that an identity link should identify the same thing unambiguously. The counting procedure can borrow that principle without requiring a site to publish structured data. Schema.org identity reference definition
An official Home Box UAE Instagram post provides an example of scope. In the earlier verification, the UAE retail site linked the regional account and the exact post was checked. The account identifies the publishing brand in that market. It does not turn a mention of a sibling brand into a Home Box mention or establish the legal name of the company behind a particular store. Use the account only for the identity decision it can support. Official Home Box UAE Instagram post
Record unresolved cases and grouped entries consistently
When a short name could refer to two firms, capture the surrounding words. A location, sector or website can make the identity clear. If the evidence remains mixed, label the occurrence unresolved and report that count. Avoid deciding from the match that would produce the most convenient result. If the match is later resolved, record the reason and the reviewer, and decide whether to update the historical series.
For a grouped list entry, separate the formatting from the entities it names. One heading can cover several firms. Conversely, several repeated headings or a list and map can describe the same firm. Count qualifying entities within the defined answer, not visual boxes. If the tool surface includes several independently generated answers, define the boundary of each run before merging evidence. The same business in a separate answer is a new observation.
A published Lunasol study of twelve Google AI Mode answers provides a real example. In the accounting run, five firms had individual entries and a sixth entry named six networks. Under the published entry-label rule, the answer contained eleven named businesses. In the gyms run, a business discussed only in prose did not qualify as an entry under that same rule. The nursery run included a domain-to-name match the study treated as arguable. These examples come from the study’s account of its coding, rather than a new replay or recoding of the answers. Lunasol’s published 12-run study
A different report could count prose mentions too. Record that as a different measurement and recalculate the relevant rows under the new rule. Do not quietly add a prose mention to one sector while excluding equivalent wording elsewhere. If you need to compare periods, keep the rule constant or label both versions so the reader can understand why they differ.
For a rebrand, record when the old and new names referred to the same company and how the report treats older answers. A historical answer can use the name current at the time. Keep that raw label and map it using the period-appropriate identity evidence. If the old name was taken up by a different organization, investigate the date, location and context before applying a global alias.
When an answer links through a directory or a map, decide what the link actually identifies. The host may belong to the directory while the linked listing is for the business. If the report measures direct citations to the official website, keep the directory link in a different category. If it measures entity references through third parties, document that broader rule and apply it to every tracked company.
Reproduce the deduplication with a small coded example
On 6 November 2026 a local check used nine fictional observations from three answer IDs. Seven were entry candidates, one was prose and one was a citation. The first answer contained three approved variants for Coast Studio, including Arabic text, and one entry for Harbour Workshop. A second answer contained the ambiguous label Coast, a prose reference to Coast Studio and an entry for Harbour Workshop. The third answer contained Coast Studio as an entry.
The checker preserved the original text, normalized strings for alias lookup and assigned a disposition to each row. Two Coast variants in the first answer were duplicates of its verified placement. The short label Coast in the second answer stayed unresolved. The prose reference and citation were kept as separate evidence and contributed no entry placements. Four placements remained: Coast and Harbour in the first answer, Harbour in the second and Coast in the third.
Under that example, Coast has two qualifying placements across three observed answers. Its entry appearance rate is two of three. The three aliases in the first answer contribute one placement. The unresolved label does not turn the second answer into a verified appearance. These figures are from fictional rows designed to demonstrate the rule. They are not estimates for an actual business or search product.
A simple implementation can deduplicate using a pair of keys: the answer identifier and the resolved business identifier. Run the counting rule on qualifying rows only. Keep the uncoded and excluded rows available for inspection. If a reviewer changes an identity decision, rerun the same aggregation and compare the counts. Use a test that contains at least one repeated alias, one namesake and one separate citation so the code cannot pass by counting strings alone.
Publish the count with its coverage and limitations
Use a Brand mention coding record with the answer ID, platform, date, original text, observation type, resolved entity, reviewer and disposition. Record the definition version. When calculating a rate, include the number of eligible observed answers and the number of planned runs. Report missing captures separately. An observed answer that contains no qualifying brand has a different meaning from an answer you could not inspect.
Have a second reviewer code a small overlap sample independently. Compare how they handled short names, branch scope and links to third-party directories. Resolve differences by updating the rule, then rerun the rows affected by that rule. Keep the previous record when a change alters an already published number. The report should allow someone to reconstruct the count without relying on a reviewer’s memory.
Look at the unresolved cases before presenting a close comparison. If two brands differ by a single placement and several names remain uncertain, include that fact beside the result. A simple sensitivity check can show the maximum change if a particular unresolved row were accepted, while leaving the headline calculation based on confirmed evidence. Keep the speculative count clearly labelled and avoid treating the possibility as an observation.
Finally, keep attribution separate. A mention count describes the sampled answers. It does not tell you whether a visitor saw the answer or whether it led to a call or enquiry. If an authorised analytics or CRM export is available, use it to report the outcome it actually captures and include unknown sources. This article used no private exports and made no commercial attribution claim.
Save a reconciliation that explains the path from candidates to counted rows. In the fixture, seven entry candidates became four placements after two duplicate aliases and one unresolved label were removed from the numerator. The prose and citation observations stayed outside the entry candidate total from the start. A reconciliation like this exposes a missing row, a duplicate introduced by a join or a decision that was applied only to one brand.
AI Visibility is published by Lunasol, which offers GEO and AEO services. The historical examples come from its published study, and the coding fixture uses fictional brands and observations. About AI Visibility



