Metric definitions

Every number Cheap Seats reports, with its formula and denominator, and how each maps to Peec AI, Profound, Gauge, Otterly.AI, Ahrefs, SE Ranking, Elmo and Limelit Open.

What exactly does each Cheap Seats number mean? Almost every metric below is a count of graded answers over a count of graded answers, worked out from what’s stored. No model is called to build a report. The same numbers are on the run page, in the run’s JSON (/api/runs/<id> and /runs/<id>/export.json, under Summary), in cheapseats report <run-id> -json (under summary) and in MCP get_run.

Terms used throughout:

• Graded answer: one answer from one answerer, on one track, that the grader graded. Errors and unfinished answers count nowhere.
• Choice prompt: a prompt that asks for a choice. Prompts marked intent = "info" ask for an explanation, so their answers count toward named only, never toward recommended, top pick, named first or top 3.
• Named: graded recommended or mentioned. For your brand this is always the grade, never the grader’s list of names, so your named rate and your share-of-voice row always agree.
• Products named: the grader’s list of every product an answer names, in order, folded onto your competitors’ canonical names. Each product counts at most once per answer.

Your rates

Each comes with a 95% Wilson interval, so 0 of 10 and 0 of 1,000 don’t read the same. For h hits of n answers, with p = h/n and z = 1.96:

centre = (p + z²/2n) / (1 + z²/n)

half-width = z · √(p(1 − p)/n + z²/4n²) / (1 + z²/n)

The run page shows these per answerer and track, and for the run as a whole. The run’s rates pool its answers: a prompt with more answers counts for more. Rankings and trends instead take the mean over prompts, each prompt counting once (see Trends).

Unbranded and branded

A branded prompt names your brand or one of its aliases as a whole word (“Is Acme any good?”). Being named in its answers proves little. The track cards lead with the unbranded rates, the same three rates over the answers to every other prompt (unbranded in JSON), and show the rates with branded prompts underneath. The prompt list marks each branded prompt. Rankings and trends use the overall rates.

Share of voice and position

The share-of-voice board lists every product the answers named, not only the competitors you configured.

• Named rate was called share of voice. In earlier versions the board’s one rate column was labelled share of voice. That number is now named rate; share of voice is the true share, which sums to 100%. The JSON still writes the old field, share, with the named rate in it, so existing readers keep working.
• Place. Your brand’s place is the position from its grade. Another product’s place is where it comes in the grader’s list of names. An answer that names your brand without listing it has no known place for it, and is left out of avg place, first, top 3 and weighted.
• First and top 3 carry 95% Wilson intervals in JSON.
• Weighted is position-weighted visibility as SE Ranking defines it: first of four scores 1, fourth of four 0.25, absent 0.
• Rank is your place on the board: 1 + the number of products named in more answers than you. Products named in the same number of answers share a rank, and a brand no answer named has none (0 in the API).

Head to head

For each product on the board, against your brand, over every graded answer (head_to_head):

Objections

The concerns raised in answers that name your brand (objections), folded for counting: lowercase, every run of punctuation and spaces one space, so “Pricing unclear.” and “pricing, unclear” count once. Each objection counts once per answer, with the first prompt it came up on as an example.

Sentiment and themes

The grader records each answer’s tone about your brand (positive, neutral, mixed or negative; none when absent) and the praise it gives, the strengths it credits you with. See How grading works.

Grades from before tone was graded have no tone, so net sentiment is over the answers that do, and the page says how many name the brand in all. With no toned answer, net sentiment is n/a.

Themes come from one more grader call after a run is graded. It groups the run’s objections, praise and reasons for picking something else into at most eight named themes, each with a kind (strength, weakness or competitive), the answers that said it, the product it helps (always one of your competitors or a product the answers named), and a quote copied word for word from what it groups. Counts, quotes and beneficiaries are checked against the inputs. With fewer than five distinct strings to group there’s no call, and the run says so.

Sources

Every page an answer cited, or its searches returned, is sorted into a kind of site:

Your sites come first, then competitors’ domains, then [sources] in cheapseats.toml, then the built-in lists; within a list the most specific entry wins. See Configuration.

For each page, host and kind:

A host counts each answer once, however many of its pages the answer used. Recall answers have no tools, so they’re left out of these rates rather than diluting them.

Buckets, outreach and found but not cited

Each host that isn’t yours gets a bucket from the products named in the answers that cite it:

With competitors configured, only they count as competitors here; with none, every other product does. A host whose citing answers name neither gets no bucket.

• Cited without your brand (outreach): the missing hosts that aren’t a competitor’s own, ranked by distinct competitors named × answers citing the host. These are the pages to get your brand onto.
• Found but not cited (found_not_cited): your own pages that searches returned and no answer linked.

Page types

The kind of site says whose page it is; the page type says what the page does, from its URL and the title the search result or citation gave it (page_type on each page, rolled up in page_types with cited, retrieved, pages and share of citations). The first rule that matches wins:

What the page is beats where it’s filed: /blog/notion-vs-obsidian is a comparison and /blog/best-note-apps a listicle. MCP get_run carries page_types and each top page’s page_type.

Citation stability

How much the hosts a set’s answers cite move from one run to the next, per track, over the same daily runs the trends use (each day’s newest finished run, the last 30 days by default). A run’s citations are a vector of host citation shares s_h: a host’s citations over all the run’s citations, so Σ s_h = 1. For each run against the one before it, with shares a and b:

• One or two transitions can be a fluke, so the score waits for three.
• The chain stops at a change of method: the run that changed it is the oldest compared.
• A run that cited nothing (an outage, or the recall track, which has no search) is skipped rather than read as every host leaving.
• A change of prompts flags its transition prompts_changed: some of its churn is the questions.

A high score says the answers lean on the same sources run after run, so a host entering or leaving means something. A low one says each run’s sources table is a sample. It’s on the set page, at GET /api/sets/{id}/citation-stability and in MCP get_citation_stability.

Trends and outcomes

var_k is a prompt’s sampling variance, p̃(1 − p̃)/n with p̃ = (h + 0.5)/(n + 1). How big a change Cheap Seats can detect explains these intervals and tests, with a worked example.

How these compare to other tools

Most AI visibility tools use the same few words for different formulas. The table below is from each vendor’s published documentation as of October 2026: Peec AI, Profound, Gauge, Otterly.AI, Ahrefs Brand Radar, SE Ranking, and the two open-source trackers, Elmo (visibility, sentiment, share of voice code) and Limelit Open (methodology). Check their pages for changes since.

The Semrush AI Toolkit sorts sources into the same four buckets Cheap Seats uses: missing, shared, strong and unique.

Migrating between Cheap Seats and these tools

• Cheap Seats’ visibility is stricter than Profound’s. Profound leaves out the runs that name no brand at all; Cheap Seats’ named rate counts every graded answer. On the same answers, Cheap Seats’ number is lower or equal. To reproduce Profound’s, take Cheap Seats’ export and drop the answers whose named list is empty.
• Profound averages per model. It works each metric out per AI model, then averages the models with equal weight. Cheap Seats’ run-level rates pool the answers; its per-answerer rates are the per-model numbers, so average those to match.
• Gauge’s share of voice depends on whom you compare with. It divides by your mentions plus the selected competitors’. Cheap Seats divides by every product the answers named, so adding a competitor to your config doesn’t change your share, except where it folds two spellings into one name.
• Peec AI’s visibility is Cheap Seats’ named rate. Responses mentioning you over all responses. Peec’s position also counts every brand detected, as Cheap Seats’ avg place does.
• Otterly.AI counts prompts, not answers. Brand coverage is the share of prompts that mention you. In Cheap Seats, count the prompts in the prompt list whose named rate is above zero.
• Ahrefs weights by search volume. Brand Radar’s AI share of voice is a share of estimated impressions. Cheap Seats weights every prompt you choose to ask the same, and estimates no traffic.
• SE Ranking’s visibility is Cheap Seats’ weighted column, the same formula.
• History. Profound applies today’s topics, tags and brands to past answers, so editing your setup changes past numbers (why past numbers change). In Cheap Seats, a run’s grades, rates and share of voice are fixed when it’s graded; a change in how answers are asked or graded starts a new method on the trend rather than rewriting the old one. To carry the trend on after a grading change, re-grade the stored answers: it’s a step you take, it keeps the replaced grades, and the run page says so. Only which pages count as yours or a competitor’s, and the source kinds, are read from today’s config for every run.
• Elmo’s share of voice also depends on whom you track. Like Gauge, it divides by your mentions plus your tracked competitors’. Its position is your rank among the tracked companies an answer names, where Cheap Seats’ avg place counts every product named.
• Elmo leaves out “I don’t know that brand”. A model reads each name match, and an answer saying it has no information about your brand isn’t counted. Cheap Seats counts it as named, graded mentioned, with the objection recorded, so Cheap Seats’ named rate can be higher on the same answers. Elmo’s visibility also counts each prompt’s latest answers, where Cheap Seats’ run rates pool the run’s.
• Limelit Open counts strings, and occurrences. Its visibility is a deterministic text search, so “I could not find Acme” counts as a mention, as its methodology says. Its share of voice counts every occurrence, not one per answer: an answer that names you once and a rival five times is a 1:5 share, where in Cheap Seats it’s 1:1.
• History, again. Elmo re-judges past answers when you add a competitor or change a name, 90 days back by default, and leaves older answers as they were, so a long chart can step at that point. Limelit Open re-reads every stored answer whenever the brands or its matching rules change.
• Your prompts move with you. Cheap Seats’ prompt sets are plain TOML in your own repo. Each prompt is a line of text, so you can paste them into any of these tools. See Compare.