Concepts

Recomma uses a small vocabulary and uses it precisely. Most of these words appear in the app beside a ? that gives the same definition — this page is the whole set in one place.

The things being measured

Project

One brand’s tracking setup: its domain, its market and language, the prompts asked for it, the models they are asked on, and the rivals it is measured against. The switcher at the top of the rail moves between them, and every figure you see is scoped to the one that is open.

A project is not a brand. It is about a brand — and it also holds the competitors, which are brands too, which is why the two words cannot be the same one. Your plan caps how many projects you run and pools the prompt budget across all of them.

Brand

A company that can be named in a chat. The one a project is about is the tracked brand; the rest are competitors. A brand is matched by name and by its aliases, so “Ahrefs”, “ahrefs.com” and “Ahrefs Site Explorer” all count as one.

The Brands page is where the set is kept — your rivals and your own row among them — because that set is the denominator of two of the four metrics.

Competitor

A rival brand you have chosen to measure yourself against. Share of voice and position are both computed across the tracked set, so the list you keep decides what those two numbers mean. A rival you have not added does not dilute your share.

Prompt

One question, tracked over time. It runs on every enabled model every cycle, and each run produces one chat. A prompt carries a topic and a funnel stage.

Topic

The subject a prompt belongs to, so a group of questions can be read together. A prompt carries exactly one topic, and the rail on the Prompts page lists whatever topics the prompts carry — there is no separate list of topics to maintain, and a topic with no prompts under it does not exist.

Funnel stage

Where the question sits in a buying decision — awareness asks what exists, consideration compares options, decision is ready to choose. Assigned when the question is written.

Chat

One model’s reply to one prompt on one day, stored whole: the text, the brands it named in the order it named them, and every source it cited. Every figure in the product is counted from these, and every figure can be opened to read them.

Mentions

The brands a model named in its chat, in the order it named them. A mention is the brand appearing in the text — not a link to it. A brand can be mentioned without being cited and cited without being mentioned.

What a reading rests on

Sample size

The chats a figure was measured over — prompts × models × sampling days in the window. Below twenty, a percentage moves too much on one answer to be reported as a reading, and Recomma greys it out.

Confidence interval

The range the true figure is likely to sit in, given how many chats it was measured over. A reading of 50% from ten chats and one from a thousand are not the same claim, and this is the difference.

Fidelity

How the chat was obtained: from the model’s own interface, the way a person sees it, or from the model’s API. Interface chats are the truer measurement; API chats are a fallback and are marked as such.

The market behind the prompts

Keyword pool

The set of keywords a category is actually contested on: everything the tracked brand or one of its competitors ranks in the top twenty for, minus brand terms, kept only where two or more of those domains rank and the intent is commercial or transactional. Rebuilt monthly and kept as a numbered version, so last month’s figure can still be reproduced. The whole recipe, and what it deliberately throws away, is on Market value.

A pool is small — a few dozen keywords is normal. It is the commercial head of a category, not a keyword export.

Intent

What somebody typing a keyword wants. The pool holds two kinds: commercial compares options, transactional is ready to buy. Informational keywords are left out, because they are answered by a paragraph naming nobody, which measures nothing.

This is the keyword’s intent, from the search data. It is a different axis from a prompt’s funnel stage, which is a property of a question you chose to write.

Rivals

How many of the tracked domains — yours and your competitors’ — rank in the top twenty for a keyword. Two is the floor for entering the pool at all: a keyword only one site ranks for is that site’s content marketing rather than the category’s.

Keyword difficulty

How hard the top ten is to enter for a keyword, 0 to 100, from the strength of whoever already holds it. Banded by colour, because nobody acts differently on 43 and 46: under 30 is a page, over 60 is a project. It weights no ranking in the product — value alone orders the pool.

Category value

What the market pays, per month, to own the intents in the pool: each keyword’s monthly searches times the share of clicks a top result takes times what advertisers bid on it. A valuation of the intent, not a forecast of revenue — being named in an answer is not a click.

Prompt coverage

The share of category value that any tracked topic reaches. The remainder is uncovered: real demand that no question measures, so the brand is scored on it in neither direction. It is the figure on Keywords with a button attached.

Captured

The share of covered value the engines actually name the brand in — each topic’s visibility weighted by what that topic is worth, rather than each question counted once. A brand doing well on cheap questions and badly on the expensive one reads as doing badly here, which is the point.

Sources

Source

A page a model reached for while answering. Recomma records both the pages a chat cited and, where the model reports them, the ones it read and did not cite.

Citations

Every appearance of a source across the chats, counted in full. One chat citing three pages of the same site counts three — which is why this is larger than the chats that used it.

Cited in

Chats that linked to a source in what they said. A chat citing it five times still counts once.

Used

The share of chats that cited a source at least once. The denominator is every chat in the window, not only the ones it appeared in.

Per chat

Citations divided by the chats that used it — how heavily a chat leans on a source once it reaches for it. Two means a typical chat quotes it twice.

Overlooked

Chats where the model read a source but did not cite it. Read-and-passed-over is not the same as never found, and the difference is what tells you whether the page or its ranking is the problem.

With you

Of the chats that cited a source, the share that also named you. A high number means the source is friendly ground; a low one means it is being trusted while you are absent from it.

Retrieved, and cited

A model can open a page and quote nothing from it, and the two are recorded separately.

The gap between retrieved and cited is itself a reading. A source opened in a third of chats and quoted in a twentieth is being picked up and put down, which is a content problem; one that is never opened is a discovery problem. The two need opposite fixes.

Source type

What kind of site the model leaned on:

The type is not decoration: it decides how much a citation is worth when opportunities are scored. Editorial coverage is weighted well above a corporate page, because it is harder to get and the models lean on it more.

Institutional was split out of reference, because the two need opposite advice. A review directory is a listing you go and claim; a .edu paper or a standards document citing your subject is ground nobody buys their way onto, and being absent from it is a research problem rather than a marketing one. It is decided by public suffix — .gov, .edu, .ac.uk and the named archives — so no classification pass is ever spent on it.

Unclassified is the one that is not a finding. It means nothing has placed the domain yet, which is a statement about us and not about the site — so it carries the lowest weight of any type, and moves to its real one as soon as the classifier gets to it. It is kept distinct from Corporate deliberately: “we have not looked at this” and “this is a company’s marketing page” are different facts, and a category holding both would be unreadable.

Gap

A domain the chats cited without ever naming you in the same chat. It is being trusted on your subject while you are absent from it, which is the cheapest place to go and be present. Gap strength is your overall visibility minus that source’s coverage of you — so a source that already mentions you as often as the web at large is not a gap, however often it is cited.

Crawlability

What a site’s robots.txt says about AI crawlers — whether GPTBot, ClaudeBot and the rest are allowed to read it. Blocking them is a choice a site makes about being quotable.

The four brand metrics

Each has its own page; in one line each:

A fifth page, Market value, covers the figures that say what the chats those four are counted from are worth.