HubSpot's AI Search Grader: What the Free Score Measures

By Muntazir · Published 16 October 2026

Type a brand name into HubSpot’s AI Search Grader and, in under a minute, it hands back a number out of 100, a breakdown across five labeled categories, and a line about whether GPT-5.4 mini, Perplexity and Gemini think of the brand as a “Leader” or a “Niche Player.” Nothing on the results page says how many times it asked, or what happens if the three models don’t agree.

That gap is the whole story. A free score that looks precise, built from a process the reader never sees, is exactly the kind of result worth checking rather than trusting.

What does the AI Search Grader actually check?

HubSpot’s own page for the tool lists five scored dimensions that sum to 100 points: Sentiment Results (40), Presence Quality (20), Brand Recognition (20), Share of Voice (10) and Market Competition (10). Sentiment carries the most weight by a wide margin — four times the weight of either of the two smallest categories. Presence Quality breaks down into mention depth, source quality and data richness; Market Competition classifies the brand as a Leader, Challenger or Niche Player and separately as an Innovator, Disruptor or Traditionalist.

The method behind all five, in HubSpot’s own words, is sending the brand’s name to three models and asking what they already know or can retrieve about it. That’s a real signal — it’s a record of how a handful of widely used assistants currently talk about a brand — but it’s a record of model opinion, not a technical audit of the brand’s own website.

How does that turn into one number?

By asking three models the same question and scoring the answers against a rubric HubSpot doesn’t publish in detail. The page states the point allocations above; it doesn’t state how many prompts are sent per model, how ties between a positive and a neutral answer are resolved, or how “Share of Voice” counts a mention that names a brand without recommending it over a competitor. A score with public weights and a private scoring process is only half-transparent — the reader can see what counts, not how a given answer became a given number of points.

Why might the same brand score differently next week?

Because nothing about the method is a lookup against a stored record — each run asks the models fresh, and language models are probabilistic. An independent breakdown of the tool that tracked this directly found scores can move between runs with nothing having changed on the graded site at all, and recommends reading the pattern across several runs rather than reacting to any single score. The same review makes a second point worth keeping: the grader reports a symptom — a weak sentiment score, a low presence number — without diagnosing why, which the reviewer says “usually lives in your content, your structured data, or your simple absence from the topic.” Finding out which of those it is needs a different kind of check than asking a model how it feels about a brand — closer to what a GEO audit looks for than to a sentiment score.

What does the score never touch?

The technical layer underneath the model’s opinion. Whether an AI crawler can actually fetch a site’s pages, whether its structured data validates against what’s visibly on the page, whether a block in robots.txt is quietly keeping an assistant’s retrieval step out — none of that shows up in a score built entirely from asking three models a question and grading the reply. A model can describe a brand warmly from training data it absorbed months ago while that brand’s current site is unreachable to the exact crawlers that would update that impression. The two facts don’t contradict each other; they’re just not the same fact, and a single composite number collapses them into one line.

A free technical scan built to probe this kind of gap runs its model-opinion step differently for exactly this reason: it never treats one sample from one prompt as a measurement, because a probabilistic answer taken once is a data point, not a verdict. The same scan’s free tier skips asking any model anything at all — questions about what a model has heard of a business, or who it would recommend instead, run only on a deeper, explicitly separate pass, so a free result and a model-opinion result are never quietly mixed into one number without saying which is which.

Does the name change anything?

Not the substance. As of October 2026, hubspot.com/ai-search-grader and hubspot.com/aeo-grader both load the same tool. HubSpot’s current page calls it AI Search Grader throughout and describes AEO as the name of its separate, paid tracking product — but AEO Grader still turns up in older coverage and some third-party guides describing what appears to be the identical free tool. Which name is older isn’t settled anywhere in HubSpot’s own public material, and the practical answer for a reader is the same either way: both URLs reach one free, one-time, model-opinion score.

A hypothetical score like “73 out of 100, Challenger” sounds like a verdict. It’s closer to a transcript of what three models said once, scored against a method the reader can’t fully inspect, about a question — “what does a model already believe” — that is only one layer of whether AI assistants can find and describe a business accurately. Treating it as the whole answer skips the layer underneath, where the actual mechanics of crawling, structured data and retrieval decide whether that belief ever gets refreshed.