The Salienz Score: how we measure whether AI recommends your brand

The Salienz Score is a number from 0 to 100 for how strongly AI assistants recommend a brand to buyers in its category. Most trackers count how often a brand is mentioned. This one asks what a buyer would ask: when people go to an AI for advice, how often does it point them to you, and how firmly? A score of 10 means your recommendations add up to being the correct top pick in 10 of every 100 relevant answers.

Because most trackers give the same credit to being named the best fit and being buried under 'others include'. Both count as one mention. A buyer reads them very differently, and the number that claims to describe what the buyer saw should too.

Position in the answer does not rescue the mention count either. In January 2026, SparkToro and Gumshoe published a study in which 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI, 2,961 runs in all. ChatGPT and Google's AI each returned the same list of brands less than once in 100 runs, and the same list in the same order less than once in 1,000. 12 How often a brand appeared held up far better than where it ranked.

Small samples mislead for the same reason. In our tests, two readings of the same unchanged market could sit 7.7 points apart through noise alone when each was built on 160 answers. At 8,000 answers each, the gap shrank to about 1.1 points. 3 Visibility numbers also rarely get checked against pipeline at all, and this one is built so that it can be.

Every answer earns points for what the assistant does with the brand, from 1.0 for naming it the top pick down to minus 0.5 for warning buyers away. The scale has six steps, and we apply it to every brand in the category the same way, the client included.

Wrong facts reduce the points. A minor error multiplies them by 0.8. A major error, being confused with another company, or appearing only under an old name multiplies them by 0.5. A warning always stays negative.

The grading is blind. The AI model that reads each answer is never told which brand is the client, two models grade, and a person audits a sample every month. In testing, a grader that nudged the client's answers up one step, one time in ten, produced a fake gain in 42 percent of markets. A blind grader did so in 4 percent, close to what noise alone produces.

Questions come from demand signals: search volume, Search Console and sales calls. A question is included when the brand would be a fair answer to it. Questions that already name the brand, such as 'Salienz vs X', do not count toward the score and are tracked separately, since being found by someone who already knows your name says little about being discovered.

Each question is asked many times across the assistants buyers use, in clean logged-out sessions in the client's market. That adds up to thousands of answers a month. How the question set itself is built is covered in how Salienz works.

• Questions people ask more often count more. The scale is compressed so one huge query cannot swamp everything else.
• Assistants your buyers use more count more.
• If an assistant shows no AI answer at all, which often happens with Google's AI Overviews, that is not held against you.
• If it answers and names nobody, that counts as a real zero.

The Salienz Score is Presence multiplied by Preference. Presence is how often AI names or cites you at all, which is the visibility number most reports stop at. Preference is how strongly it recommends you when it does. A brand that appears in 34 percent of answers with an average of 0.53 points scores 34 times 0.53, which is 18.

The split is what tells you what to do next. If the score falls while Presence holds, the assistants still name you and recommend you less firmly. If Presence falls while Preference holds, you are being left out of answers. The score can also drop below zero when warnings outweigh recommendations, and 100 would mean the correct top pick every time.

Every Salienz Score comes with a 95 percent range, and the same question set is run twice to measure normal noise. In testing at 8,000 answers a month, that range averaged 0.8 of a point either way, and a month-to-month move for the brand under test had to pass about 1.4 points before it stood outside normal noise. Four more checks sit behind that range.

• Net movement is measured within each assistant. Your share of recommendations is compared with the category's inside one assistant, then the results are combined, so a model update that hits every brand does not look like your win or loss. When one assistant turned cooler towards every brand in testing, the raw score raised a false alarm in 63 percent of markets and the net figure in 9 percent.
• About 10 percent of questions are held out and never shared with the content team. If tracked questions improve and held-out ones do not, the gains are not real. In our tests, content written only for the tracked questions lifted the score on them by a median of 14 points and left the held-out questions flat.
• The grading is blind, as described above.
• Every brand the assistants name is tracked, not a hand-picked competitor list.

The score has a rule for each case that a plain count gets wrong.

• Look-alike or common-word names: only the real company counts.
• Rebrands: mentions under the old name get half credit, and the score tracks how fast the new name takes over. Salienz went through this itself when it changed its name in September 2026.
• 'Best for X' answers count as a top pick only when X matches the buyer who is asking.
• Sponsored placements are excluded.
• Fifteen-brand dumps are worth little to anyone on the list.
• Categories where AI names nobody are reported as open space, not as a zero.

Because the standard metrics count appearances, and buyers act on recommendations. A mention rate is the share of answers that name you. Share of voice is your slice of all the brand mentions in those answers. Neither can tell a top pick from a passing mention, and in our tests a mention rate kept showing a gain while recommendations fell.

Some trackers add an average position. The order of a list barely repeats, as the SparkToro and Gumshoe study above found, so the Salienz Score leaves position out and grades what the answer says about you. 1

We checked the difference before relying on it. On 300 simulated markets, where a simulated buyer read every answer and chose a brand, we ran the score beside copies of a common mention-rate formula and a common sentiment composite. They are copies of the formulas, not anyone's product. 3

The gap is widest where it costs a client most. When a brand was named more often and recommended less firmly, the mention rate reported a gain every time, and the score flagged the drop in 90 percent of cases. When half of a brand's appearances came back under its old name, the score caught it every time, because old-name credit is halved by rule. A mention count has nothing to mark down.

The rest of the difference is in how the number is produced. Every score carries a 95 percent range, movement is measured inside each assistant, a tenth of the questions are held out from the content team, and the grader never knows who the client is.

Expect lower numbers than a mention rate gives you, because a mention is not a recommendation. In test markets, the middle half of category leaders scored between 35 and 45, and the middle half of mid-pack brands between 11 and 16. 3 Real categories will differ.

• Use Presence and Preference to explain why the score moved.
• Read monthly changes as directional. Judge a 90-day programme on all of its answers pooled together.
• Pipeline is still the outcome. The score is the early signal you steer by.

It is not revenue, market share, a rank tracker or a guarantee. Why analytics miss most of what an AI recommendation does to pipeline is covered in why GA4 shows almost no ChatGPT traffic, and a baseline score is part of the Foundation audit.

Four ideas in the score came from other people's work.

• SparkToro and Gumshoe's January 2026 study on recommendation variance, for the finding that how often a brand appears is stable where its rank is not. 1
• Aggarwal and colleagues, 'GEO: Generative Engine Optimization', presented at KDD 2024, which scored visibility with an AI-judged measure covering relevance, influence, the likelihood of a click and other factors. 4
• The recommendation rate reported by some AI visibility tools, which counts the share of answers that actively recommend a brand. 5
• Holdout testing, borrowed from machine learning.