TL;DR
- A source link and a brand name in the answer are different events. A page can receive credit without the prose naming its owner.
- Read the base behind a ghost-citation rate. One study's detailed results describe domain appearances, not just citations or all AI answers.
- A named brand is not always a recommended brand. Save the surrounding claim before judging the result.
- Report links, names, recommendations and business outcomes separately, using the same declared sample.
Choose the outcome before the score
For generative engine optimization (GEO), first define what your team wants to observe. Is it use of your page as a source, an explicit brand name or a recommendation for a stated need? Each can matter. Combining them into one score can hide which result the work actually produced.
What Did the Ghost-Citation Study Actually Count?
Semrush's June 2026 study, conducted with Kevin Indig and Growth Memo, recorded 3,981 domain appearances across 115 prompts, 14 countries and four AI engines. Its detailed breakdown was 61.7% source-link only, 13.2% both linked and named, and 25.1% named without a source link.
| Observed appearance type | Share of domain appearances |
|---|---|
| Source link, no brand name | 61.7% |
| Source link and brand name | 13.2% |
| Brand name, no source link | 25.1% |
Source: Semrush, June 9, 2026. Base: observed domain appearances, where a link, name or both occurred. This excludes absence and is not a share of all AI answers.
The article's headline refers to about 62% of citations. The detailed table uses all appearances as its base. Those are different sets. Use the detailed base when reusing the figures. The public methods do not give collection dates or model versions, so the shares should not be treated as a current forecast for your brand.
The result explains why a source-count report needs more detail. It does not show whether a person noticed the source badge, recognized the company or later bought from it. A link-only result is not the same as zero human exposure.
Semrush / Growth Memo, June 2026 publication. 3,981 domain appearances, 115 prompts, 14 countries. Base excludes absence; these are not shares of all AI answers or just citations. Collection dates and model versions not given. Semrush, June 9, 2026.
Does Buyer Trust Research Make Every Citation a Brand Win?
No. A survey about perceptions cannot settle what a particular answer showed. G2's 2026 AI Search Insight Report says 85% of respondents thought more highly of a vendor cited by AI. Its global online survey covered 1,076 B2B decision-makers in March 2026. This is a stated perception, not a test of link-only answers or sales impact.
| Direct research | Reported result | What the study measured | What remains untested |
|---|---|---|---|
| Semrush, June 2026 | 61.7% were source-link only | A category within observed domain appearances | Whether people recognized the brand |
| Semrush, June 2026 | 13.2% were linked and named | Two events occurring in the same record | Whether the text recommended the brand |
| G2, 2026 | 85% reported a better view of a vendor cited by AI | Buyers' stated perceptions | An effect from an unnamed source link |
G2's public discussion uses both “mentioned” and “cited” around that finding. Do not assume its wording matches a monitoring tool's exact event rules. The studies are useful together because they reveal a measurement gap. One records output features; the other asks buyers about their views. Neither provides a conversion factor from links to brand trust.
For a buyer of GEO services, this changes the proposal review. Ask what the provider means by visibility and request full answers with exact source links. A total can be accurate while its label implies a stronger result than the records support.
How Should a B2B Report Label Each Result?
Use a simple coding rule that another reviewer can repeat. The matrix below is an editorial reporting aid. Count a brand name in the answer's prose; record source-card or badge visibility separately if that matters to your goal.
| Source link to your site? | Your brand named in prose? | Label for this record | Further check |
|---|---|---|---|
| Yes | No | Cited only | Does the page support the nearby claim? |
| No | Yes | Named only | Is the name accurate, and in what context? |
| Yes | Yes | Cited and named | Is the brand actually recommended for this need? |
| No | No | Neither observed | Keep the answer in the declared sample |
Consider an illustrative answer whose prose says, “Check whether the offer includes setup and training.” Its source list links to Vendor A's guide, but the prose never says Vendor A. For Vendor A, the record is cited only. A source card may still display its name; that should be a separate field, rather than silently changing the prose-name count.
Now consider an answer that names Vendor A only to say it lacks a required feature. That is a mention, but not a recommendation for that requirement. Keep the relevant passage so the reader can see the difference. A positive-sounding summary in a spreadsheet can otherwise conceal a poor fit.
Keep prompt, engine, product mode, date and market with each saved answer. Use a fixed rule for aliases and repeated links. These details make a comparison in your AI visibility measurement plan meaningful. Pair the record with a source-support check before using it as evidence in a sales or strategy discussion.
Ghost Citations and B2B Brand Mentions: Common Questions
Should we try to eliminate ghost citations?
No. A useful research page may serve the buyer as a source even when the owner is not named in prose. Judge it against the page's purpose. If your goal is brand introduction, report explicit names as a separate result.
Does naming our company more often on the page guarantee mentions?
No. Make ownership and product names clear where they help readers, but avoid repeating the brand in every paragraph. A change in your copy needs its own output check. The studies above do not establish a guaranteed naming tactic.
Can one answer count as both cited and named?
Yes, if it meets both rules. Report the overlap so a combined total does not count the same record twice. Keep the denominator clear: answers, domain appearances and individual links are different units.
What should a report keep beyond a screenshot?
Keep the full answer, exact links, collection context and the coding decision. Note how unavailable or failed responses are handled. A screenshot can help a reviewer inspect one record; it cannot establish how often buyers encounter that result.
Review what your visibility report counts
If a visibility report mixes source links with brand recommendations, bring its counting rules and sample answers to a SemShift AI visibility audit discussion. We can discuss which labels the evidence supports and what a useful measurement scope should preserve.