What Is an AI Recommendation?
GEO and AEO platforms increasingly promise recommendations as well as citations. What do they measure—and can they show you what changes the choice?
An AI assistant can cite your website, praise your product and recommend your competitor. It can also recommend you in its first answer and abandon you when the buyer asks the next question.
An AI recommendation is an explicit endorsement of an option for a stated need. Its commercial significance depends on when it occurs, what requirements have been considered and whether it survives further questioning.
That distinction matters as GEO and AEO platforms add recommendation measures to their visibility dashboards. Some classify the strength of an endorsement. Others count outright selections. A score alone does not explain how the decision developed or what a brand should change.
A recommendation score records the verdict. A decision analysis identifies where that verdict changed, whether the change was justified and what intervention is worth testing.
Follow the recommendation beyond the first answer
In an AIVO study of US television buying scenarios, an established brand appeared in 102 of 135 opening answers. After the shopper specified a budget, that fell to 42. After a question about full-array local dimming, it fell to 29. At the final turn, five answers still named the brand, but none selected it outright.
Two competitors received outright recommendations in 120 of the 135 conversations.
| Stage of the conversation | Conversations naming the brand |
|---|---|
| Opening question | 102 |
| Budget specified | 42 |
| Product feature checked | 29 |
| Final answer | 5 |
These are aggregate counts across complete conversations, not one transcript. The study covered budgets of US$1,000 or less; every conversation explicitly asked about local dimming at turn three. The results describe that tested market and requirement, not the brand’s entire television business. [1]
An opening-presence measure would have shown a widely recognised brand. A final-selection measure would have shown a brand that won none of the tested conversations. The conversation analysis showed where the brand fell away and which claims accompanied its exclusion.
Some products genuinely lacked the feature being tested. Elsewhere, assistants used incorrect prices, misrepresented specifications or presented a discontinued range as current. Shopping-agent runs also exposed difficulty verifying price and availability. [1]
Those are different failures. They demand different interventions.
One word conceals several different events
Consider three illustrative answers:
- “Brand A is worth considering.”
- “Brand A is a good option if you need a lightweight shell.”
- “For your stated requirements, choose this Brand A jacket.”
The first offers consideration. The second makes a conditional endorsement. The third selects an option for a brief. All can influence a buyer, but they should not be reported as interchangeable outcomes.
Even the third may be provisional. The customer can introduce a delivery deadline, compatibility requirement or objection that changes the choice.
A first-answer recommendation is real at that point. It becomes evidence of a completed decision only when the relevant decision process has been tested. Four turns are not inherently sufficient, just as a long conversation is not inherently realistic. The journey must reflect the requirements and actions material to the purchase.
What recommendation platforms actually measure
MentionLab’s Recommendation Strength Rate distinguishes Chosen, Described, Named, Criticized and Dismissed, then combines graded mentions into a score. That is more informative than counting appearances. Its AI Visibility page does not describe a connected follow-up sequence that tracks a brand through to a final decision; the public description alone cannot establish whether the product supports one. [2]
Mixtape Partners makes its boundary explicit. Its Pick Rate specification counts a single recommended brand and publishes scoring, denominator and uncertainty rules. It also identifies the measure as single-turn and separates it from diagnosis and purchase outcomes. [3]
The wider industry recognises conversation depth. Profound has published research on multi-turn commercial conversations; Scrunch has examined recommendations and subsequent consumer behaviour. Those studies establish that the industry is investigating recommendation behaviour; they do not establish equivalence between commercial measurement products. [4][5]
The scrutiny belongs at the level of the claim. Classifying an endorsement can substantiate a claim about endorsement. It does not automatically substantiate sustained preference, successful remediation or additional purchases.
The displacement is where diagnosis begins
AIVO Meridian examines complete conversations to locate displacement: where a brand loses its position, which alternative gains and what stated judgment or evidence accompanies the switch.
AIVO’s AI Win-Rate counts outright selections at the decision point across tested buying conversations. Shared recommendations and no-pick outcomes are reported separately. Earlier leadership, displacement and resilience under challenge sit alongside that result rather than being blended into the win rate. [6][7]
For a person using an assistant, the important event may be an objection that changes the assistant’s advice. For an agentic journey, it may be a product check, inaccessible page or unverified condition that causes the agent to remove an option.
Neither observation automatically proves an error. An unsuitable product should lose. The diagnostic task is to distinguish a justified exclusion from a failure to retrieve, represent or apply relevant evidence.
AIVO calls one such failure a linkage gap: the system demonstrates knowledge of relevant brand facts but fails to carry those facts into its recommendation. Reasoning-path dissonance describes an inconsistency between the assistant’s stated assessment and its eventual choice.
If an assistant acknowledges that a product meets the requirements, then excludes it using a conflicting account of those same facts, the transcript exposes a problem worth investigating. It does not reveal private internal model reasoning. The analysis concerns observable answers, actions and sources.
Surgical remediation needs a traceable target
The television study illustrates why a general instruction to publish more content can miss the problem.
An incorrect retailer price calls for a correction to that evidence. An obsolete model recommendation calls for a clear account of the current range. If availability depends on location, the agent needs a usable way to supply that location and check the result. A genuine product limitation may require a different product or positioning decision.
A linkage failure requires testing how the relevant product fact is connected to the buyer’s requirement. Repeating a specification more often may accomplish little if the assistant still fails to apply it when choosing.
“Surgical” means the intervention addresses a demonstrated failure in a defined decision context. It is a testable hypothesis, not a promise to control the assistant.
The evidence chain should remain visible: conversation, displacement, diagnosis, correction, repeat test.
Rigour is in the design, not the dashboard
AIVO’s published chronology traces recommendation research to February 2026 and Agentic Brand Control to 13 May. Its September AI Win-Rate methodology formalises decision-stage measurement, scenario-set disclosure, uncertainty and before-and-after comparison requirements. [6][7]
A buyer assessing any provider, including AIVO, should ask:
- Does the test follow connected conversations or ask independent prompts?
- How were buyer scenarios selected, and were brands named in advance?
- What counts as an outright win, a conditional endorsement or a shared choice?
- What sample sizes, repeated runs and scoring checks support the result?
- Which model versions and retrieval settings were used, and when?
- How will the repeat test distinguish intervention effects from noise and model changes?
Identifying a displacement is descriptive. Explaining it requires evidence. Demonstrating that a correction improved outcomes requires a suitable comparison. A rigorous method keeps those claims separate.
Connecting the answer to the purchase
Meridian identifies the question and evidence gap associated with a lost recommendation. AIVO Assist gives the visiting agent a route to obtain an approved answer.
Assist applies product facts to the buyer’s circumstances, asks for missing details and supplies sources. It can also say that no option fits. This helps agents establish suitability while their customer is still choosing. [8]
Where a supported integration preserves the answer reference through a click, booking, basket or order, the commercial event can be matched back to the answer. That provides a route from exploration to attribution; coverage depends on the integration and the reference surviving the journey. [8]
AIVO reports 14 of 14 eligible bookings correctly attributed, with no false credits, in a separate controlled test. Merchant customer outcomes remain pending. A recorded connection does not establish that the answer caused an additional purchase; incrementality requires its own test. [8]
Recommendation, task completion, attributed purchase and incremental purchase answer different commercial questions. Each needs its own evidence.
Ask to see a lost decision
Recommendation measures are a useful advance on counting citations alone. Their value depends on the event they capture and the conclusions they can support.
Before paying for an AI recommendation score, ask to see a lost decision: the conversation, the displacement, the evidence behind it and the correction tested.
If the provider cannot show that chain, you still do not know what to change.
Sources
- AIVO, Consumer electronics brand-choice study, October 2026. 135 four-turn conversations across ChatGPT, Gemini and Perplexity, with five runs per scenario and assistant; separate shopping-agent tests. Full report reviewed for this article; brand and commissioning organisation withheld. An anonymised extract is available from AIVO.
- MentionLab: AI Visibility and Recommendation Strength Rate.
- Mixtape Partners: Pick Rate Specification, version 1.0, 26 August 2026.
- Profound: Commercial conversations in ChatGPT more than doubled in a year, 13 August 2026.
- Scrunch: The prompt-to-purchase pipeline, 17 June 2026.
- AIVO: AI Win-Rate definition and research chronology.
- Sheals and de Rosen: AI Win-Rate—A Board Metric for AI-Mediated Buying, Methodology v2.0, 8 September 2026.
- AIVO Assist: operation, attribution mechanism and controlled-test disclosures.
Disclosure: The author is co-founder and CEO of AIVO, Inc. This article assesses public descriptions, not competing platforms through a hands-on audit. AIVO evidence includes working papers and company-run studies and tests; independent validation is not implied. Sources checked 19 November 2026.
Comments ()