What Is Share of Recommendation in AI Search, and How Do You Measure It?

Most AI visibility reports answer one question: did the engine name your brand? That misses the part buyers act on. An answer can list five brands, recommend two, call a third a budget pick, and mention yours only as the one to avoid if you need fast support, and a presence metric still scores it as a win.

Share of recommendation is the share of relevant AI answers that put your brand forward as a choice the buyer should consider or pick. It sits one level above mentions and citations, and it is the closest thing AI search has to a shortlist metric. This post covers what counts as a recommendation, how to measure share of recommendation without fooling yourself, and what actually moves it.

Mentions, citations, and recommendations are different events

Most AI visibility reporting starts with presence: did the answer contain your name? That is useful, but an answer can mention a brand in several ways, and only some of them help you.

• A mention means your name appears anywhere in the answer, including as a comparison point or a warning.
• A citation means one of your pages is linked as a source. You can be cited for a definition without being suggested to anyone.
• A recommendation means the answer presents you as suitable for the buyer's need, as a top option, or as the pick.

The industry is starting to separate these. The IAB's August 2026 guidelines for measuring AI visibility group metrics into presence, prominence, portrayal, and persuasion, and list recommendation strength under persuasion, the layer that asks whether visibility drives action. Share of recommendation is a practical way to track that layer.

How to measure share of recommendation

The formula is simple. The discipline is in the inputs.

1. Choose eligible prompts. Use questions where a recommendation is possible: "best X for Y", "which should I choose", "alternatives to", "is X worth it". Informational prompts ("what is X") rarely produce a pick and will dilute the number.
2. Write the threshold once. Decide in writing what counts as a recommendation, and what does not: an example, a comparison point, or an option the answer advises against. Apply the same rubric to every answer, every week.
3. Run each prompt many times on each engine. AI answers change from run to run, so one response is a sample. SparkToro's research found that brand lists almost never repeat, and suggests 60 to 100 runs of a prompt to see a model's real set of recommendations.
4. Calculate per engine and per topic. Share of recommendation equals answers that recommend you, divided by eligible answers, expressed as a percentage. Keep the engines separate: a blended figure hides an engine where you are missing entirely.
5. Add strength as a second layer. Being the lead pick is different from being one of six options, so track how often you are the first recommendation as well.

In this illustrative example, the brand is visible in 35% of answers but recommended in 15%, so fewer than half of its mentions turn into a recommendation. That gap is the work.

What moves your share of recommendation

AI answers recommend brands they can justify for a specific need. Engines assemble answers from pages, threads, and reviews they trust, so the fixes are concrete:

• Pages that answer one buying question clearly, such as a use-case or comparison page, with the facts an answer needs to say "this one fits"
• Accurate, consistent facts about your products, positioning, and policies across your site and the sources AI reads, with structured data that helps AI engines read each page
• Third-party proof on the sources engines lean on: review sites, community threads, and publications
• Corrections where an engine describes you wrongly, because a recommendation built on an outdated claim helps no one

This is the core of our GEO service: the sources AI cites instead of you become content, community, and placement tasks, and the effect is measured per engine after the work ships.

Where share of recommendation misleads

A clean formula can still produce a misleading number.

• Small prompt sets. The IAB guidelines separate directional data from decision-grade data, judged on query volume, sample size, prompt type coverage, testing cadence, and reproducibility. A handful of prompts is directional at best.
• A moving threshold. If the definition of a recommendation changes, the trend line breaks. Version your rubric and note every change.
• The wrong need. Being recommended as the cheap option when you sell premium raises the number and hurts the brand. Read how you are recommended, not just whether.
• Blended engines. A brand can be the lead pick in one engine and absent from another, and an average across engines hides both facts. Report each engine on its own line.
• No link to revenue. Share of recommendation is a leading indicator, so pair it with AI referral sessions and revenue.

For one clean color cosmetics brand, becoming the most cited name in AI answers for its category, with a 33.28% share of voice, came alongside 56% growth in AI-driven revenue year over year. Presence and revenue read together tell the story; either alone does not.

Count the answers that choose you

Mentions tell you that AI engines know your brand. Share of recommendation tells you whether they would put you on a buyer's shortlist, and why not when they do not. Define the threshold, run enough prompts often enough, and read the answers behind the number.

Curious how often the major AI engines recommend your brand, and which sources they use to justify it? Talk to our team and we will show you your category, engine by engine.