SoleStats asks AI assistants the questions runners ask before buying shoes, many times over, and scores what comes back: which brands get recommended, whether the facts are right, what gets cited, and where the sale goes. It is an independent project. Scores describe AI answers on the dates sampled, not the products and not the brand. Brands get their numbers a few days before each edition.
Definitions
The same text appears when you hover or tap a column or metric name on the data pages.
Reading the tables
Every number is a rate across repeated answers, never a single answer. The ± beside a score is how much it moved between repeats of the same prompts; gaps under about five points are noise. The brand table sorts by the AI score; a dash means the subject has no data for that brand in this edition, and "not enough data yet" means the brand has too few scored subjects for an AI score. The "Where the sale goes" table is the share of answers in each category that point to each place to buy. The engine cards show how many answers each engine gave, how many ran a web search, how many named their sources, how many searched without naming any, and how many answered without searching.
A yellow banner marks a preview: a partial or sample round whose numbers are not findings.
Downloads
Every prompt, every answer with its citations, every detected mention and route, and the scored table are free to download as CSV from the front page. Raw API responses are kept but not published.
What we ask
138 prompts written the way runners type them: 102 about shoes (grouped by need: starting out, budget, cushion, stability and injuries, fit, racing, trail, drop, conditions, specific people), 28 head-to-head questions between the eight biggest shoe brands, and 8 about where to buy. Outside the head-to-head set, no prompt names a brand. The full list is in the downloads.
Who we ask
Three AI assistants through their APIs with web search turned on, so answers resemble the consumer apps and carry citations. Each prompt is asked three times per engine. A full round runs every month. Accurate's spec questions and the hand checks behind Accurate and Shoppable run once a quarter; in the months between, the scores use the latest checks, and each brand page says which round they come from. The front page names the engines and model ids in the current edition; an engine marked pending is being added and its answers join the same edition when they are in. We never present a single answer as a finding; everything is a rate across repeats, with a margin that shows how much it moved.
How shoes are scored
- Recommended (35%): how often the brand is named on prompts where it has a credible shoe for that need, plus its head-to-head record, how often it is the first pick, and how consistent that is across repeats.
- Accurate (25%): seven facts per model (current version, price, weight, drop, stack band, widths, category) checked against the brand's own US site on a recorded date. "Not sure" scores the same as wrong.
- Shoppable by AI (20%): the share of a twelve-item checklist the brand's US site passes, checked by hand once per round: AI crawlers allowed in robots.txt; schema.org Product data with offers; price, widths, stock and specs readable without scripts; the current version named clearly; returns and shipping findable within two links; crawlable pages for stability, neutral, trail, racing and wide; a path to cart and checkout without a login wall or bot block; a store locator. The full list with how each item is checked is in the downloads as the checklist CSV.
- Sourced (10%): what AI cites when it talks about the brand: the brand's site, established review sites, forums, retailers or junk. Only sources an answer names count. Two kinds of answer are counted beside the score instead: source not provided, where the engine searched the web but named none of what it found (Claude's current search tool works this way), and no search, where it answered from what it already knew (Gemini decides for itself whether to search, and mostly does not). Both are real ways shoppers get answers, so they still count for Recommended and Where the sale goes. The pages a search returned are in the downloads. Another brand's site does not count in its favour. A brand's own sites include its outlet stores and its sites outside the US, which count the same and are shown separately. With fewer than five answers that name a source, Sourced is not scored and the AI score is weighted over the other subjects.
- Where the sale goes (10%): whether answers that name the brand send the shopper to the brand's site or run specialty, or elsewhere. Only the brand's own site counts as its site: a link to another brand's site is not a place to buy this brand's shoes and is left out. A general line such as "buy from the brand's own website" counts for every brand the answer names.
The AI score is one number out of 100 for how AI assistants treat a brand: the weighted mean of the subjects above that the brand has scores for, with the weights rescaled to add up to 100; it needs at least three. Until Accurate and Shoppable are scored, the AI score's definition (tap the column heading) says which subjects it covers in that edition and which are still to come; until then it mostly reflects Recommended, so compare brands on the subject scores as well.
Scores are 0 to 100 with no curve. Eighteen brands are watched: the eight biggest plus ten more scored on the prompts where they have a credible shoe, as recorded in the verified spec sheet. A brand with no verified models yet is scored on every shoe prompt, and its card says so.
Brands by runner need
The shoe questions fall into eleven groups by what the runner needs: starting out, budget, cushion, stability, fit and width, racing, trail, drop and minimal shoes, weather and surface, who is running (women, older runners, kids and so on) and questions about brands in general. The heat map on the front page, and the row on each brand page, shows for each group the share of answers that name the brand, counting it once per answer; the tip on each cell adds the share where it was the top pick and how much the share moved between repeats. Every brand is counted on every group, so a trail brand shows low on racing by design: read a row for where a brand is and is not part of the conversation, not as a grade. The colour has six fixed steps so editions compare; the number is always printed. It is not part of the AI score.
Head to head
Twenty-eight questions ask an engine to choose between two of the eight biggest brands for a daily running shoe ("Brooks or Hoka: which should I buy, and why?"), with the order of the two swapped between repeats. Each answer is read for its verdict: the brand it says it would pick, a tie when it explicitly calls it even or says it depends on the runner, or unclear when it lays out both without choosing. The head-to-head page shows, for every pair, the share of answers that favour the row brand, with ties counted as half and unclear answers left out; the tip and the table give the counts and the split by engine. Each pair has nine answers (three engines, three repeats), so only clear leans mean anything. The same record is one part of Recommended.
Where the sale goes, by brand
Each brand page splits the places to buy named in answers that mention the brand into five kinds: the brand's own site (outlets and its sites outside the US included), run specialty stores and their online shops (Fleet Feet, Running Warehouse, a local running store and so on), Amazon, big box stores (Dick's, Walmart, Target), and everything else (REI, Zappos and the like, listed under the bars). A place counts once per answer, and a link to another brand's site is left out, as in Sale goes home, which is the first two kinds added together.
Sources to earn
Each brand page lists the ten sites AI names most often as its sources in answers that mention the brand, and the sites it names in answers about competitors but never in an answer about the brand. A gap has to come from a group of questions where the brand also appears (a road brand is not shown trail sites it was never asked about), be named in at least two answers in that group, and not be another brand's own site; gaps are ranked by how many answers named them. The top three that are review sites, forums, news or retailers become Top fixes under Sourced; medical pages and other companies' stores are listed but not suggested. Only sources an answer names count, so in this edition these lists come almost entirely from ChatGPT, which names its sources, with a few from Gemini when it searched; Claude's search tool does not name the pages it used. The card says how many answers each list rests on.
By engine
Under each score on a brand page, the same score is worked out on one engine's answers at a time, for the subjects that come from the answers (Recommended, Sourced and Sale goes home; Accurate and Shoppable come from the hand checks). Where one engine is more than ten points below every other engine that has a score, with at least twenty observations behind each, the engine is marked and the gap becomes a Top fix for that subject. Sourced has no score on an engine that names no sources, and a score held back for too few observations takes no part. When the same engine is that far below on the same subject for more than half of the brands compared, the gap says more about the engine than about any brand: it is marked on each brand page but is not a fix, and the edition summary says it once.
This edition
The edition summary near the bottom of the front page is written by the site from the edition's numbers, with no hand-written text, so each new round rewrites it: the brand with the highest AI score and the one with the largest share of voice (a gap under five points is called level), the widest voice-vs-size gaps among brands with a brand or running-level revenue figure, the subject and brand where the engines' scores are furthest apart (engines with at least twenty observations), and how many answers named their sources, by engine, with the site outside brand sites named most often.
Models named
Answers are read for the shoes on the spec sheet, by name and by the short names people use (Nimbus for Gel-Nimbus, 1080 for the Fresh Foam X 1080), with any version number written after them. A name that is also a common word (Ride, Guide, Launch, Boston) counts only with its brand's name just before it or a version number after it. Each brand page lists the brand's shoes named in answers that mention the brand, the share of those answers naming each, the versions named, and the other brand's shoe named most often beside it; the head-to-head page shows the shoe of each brand named most for each pair. The list of names and short names is kept by hand.
Version and price
Two numbers on each brand page say whether AI points shoppers to the shoe a brand sells today. Current version: of the times the brand's shoes are named with a version number, the share naming the version on sale now, taken from the spec sheet once each shoe's current version has been checked by hand; until then it reads "waiting for the spec sheet". Steered to older or cheaper: the share of answers naming the brand where the sentence or bullet about the brand or its shoe points to last year's model or an older version, or to a sale, discount, outlet or clearance, from short phrase lists, with the same share across all brands beside it and the split by engine in the tip. Both are context for a brand team, not part of the AI score.
When not you, who
For each brand, the shoe questions where it has a credible shoe (one of its spec-sheet shoes is made for that need) are read for the answers that leave the brand out, and for which brand those answers picked first. The brand page lists the five brands picked most often in its place, and the need groups where it was left out most, with the brand picked first there. Until a brand's spec-sheet rows are checked by hand, its draft rows decide what counts as credible, and the card says so.
What AI says about it
Each brand page shows how often the answers that name the brand describe it in a given way: cushioning, stability, durability, price and value, width and fit, speed, light weight, and suitability for heavier runners, and how often they warn about it: fit warnings (runs narrow, size up), durability complaints, ride complaints (too firm, unstable) and "avoid" or "not for" language. A word counts only in the sentence or bullet that names the brand or one of its shoes (a short shoe title carries to the bullets under it), and a warning right after "not", "no" or "without" does not count. The words are a short list checked by hand, with no AI reading involved; the same share across all brands sits beside each number, and a difference of ten points or more is marked.
The brand page
Each brand page opens with what a brand team needs first: the AI score with each subject on one line and its score on each engine (a ▼ marks an engine far below the others), what to fix, which brands AI picks when it leaves the brand out, and any further sources to earn beyond those the fixes already name (review sites, forums, news and retailers only; medical pages, other companies' stores and file hosts are never suggested). Every other view sits below under Detail, one tap to open: by runner need, head to head, models named, what AI says about it, version and price, where the sale goes, the full source lists, revenue, the top 10 at the majors, sample answers, and how the scores are built.
Revenue
Each brand page shows the latest revenue the brand, or the company that owns it, has reported publicly: from its 10-K, 20-F, annual report or official results release, led by its year-over-year change in the currency it reports in (where a company also gives a currency-neutral rate, it is in the notes), with the amount and a link to the source. Where the owner does not break the brand out, the figure is labelled as the footwear, segment or whole-company number it is. Private companies are shown only with figures they have announced themselves. Revenue is context for the size of each brand; it is not scored, and no link between revenue and the AI scores is claimed. The figures are refreshed once a quarter and are in the downloads.
Share of voice
The share of voice page puts each brand's share of AI answers beside the size of its business. Share of voice is the brand's share of all brand mentions in answers to the shoe questions, counting each brand once per answer and leaving out head-to-head and where-to-buy questions; it comes with the share of top picks, the share of answers that name the brand, the split by engine and a margin from the repeats. Running-shoe market share is not published, so the first chart sets share of voice against each brand's narrowest reported revenue figure instead, converted to US dollars at the Federal Reserve's yearly average rate only to put brands on one scale (the Revenue card keeps the company's own currency). Filled dots are the brand or its running business; rings are wider figures, all footwear, a segment or the whole company, and overstate the running business. The second chart sets the change in share of voice since the previous edition against reported revenue growth; it fills in from the second edition. Share of voice is not part of the AI score.
Voice vs size divides a brand's share of voice by its share of the revenue reported by the brands that have both numbers, in US dollars. At 1× AI mentions the brand in line with its size; above 1×, more; below 1×, less. The revenue figures are not all at the same level: where a company reports only its footwear, a segment or its whole business, the figure overstates the running brand, so its ratio reads low, and the table marks it with a ring. It is context for a brand team, not a cause in either direction: AI mentions do not drive revenue on this evidence, and revenue does not explain the mentions.
Top 10 at the majors
For the top ten finishers of the men's and women's races at the World Marathon Majors (Tokyo, Boston, London, Sydney, Berlin, Chicago, New York) and the World Championships marathon since 2025, we record the shoe brand each raced in: the race-day shoe where running press, a race photo or the brand's own release reports it, otherwise the athlete's sponsor, and the brand page says which. Places and times come from the official results. A shoe brand counts only once its source has been read; places whose shoe is not confirmed are left out of the shares. The third chart on the share of voice page sets each brand's share of voice against its share of those top-10 places. Race results are context and are not scored. Wheelchair races are not included.
What it can't tell you
AI answers change week to week and engine to engine. A score describes the answers we sampled on the dates given, not the products and not the brand. Eighteen brands and one round cannot support any claim about sales; we show search interest and public scale figures beside the scores only as context. Detection of brand names, citations and retailers is automated and then reviewed by hand; the raw answers are downloadable so anyone can check.
Run it on your own category
The prompt templates, the rubric, the runner script and the scoring code are open. Swap the brand list and the prompts and the same pipeline scores any category.