Supracited

Sample audit

A low-profile keyboard brand, audited

Demonstration audit on a real small brand, prepared without its involvement, Claude route, 50 category answers, October 2026.

Asked ten keyboard questions that name no brand, the engine named the brand in 5 of 50 answers: between 0 and 28 in every 100, 7th of the 15 brands named. Keychron was named in 45 of 50, NuPhy in 31 and Logitech in 26.

Sampling record: 12 questions (10 that name no brand, 2 that name the brand), 5 repeats each, 60 of 60 answers collected on 2026-10-16 between 19:39 and 19:45 Pacific time, 0 failed. Engine: Claude with web search, through the Claude Code CLI research route, not the consumer app. The brand's name and its product names are withheld on this page.

Share of answers

How often each brand is named

Engine in this table: Claude with web search.

The same 50 answers to ten questions that name no brand, so every rate compares directly. The filled block spans the 95% range; the tick marks the share observed.

Bar: the 95% range. Tick: the share observed.

  1. Keychron 74 to 100% · 45 of 50
  2. NuPhy 36 to 88% · 31 of 50
  3. Logitech 22 to 80% · 26 of 50
  4. Royal Kludge 6 to 54% · 14 of 50
  5. Corsair 0 to 38% · 8 of 50
  6. Akko 2 to 26% · 6 of 50
  7. The brand 0 to 28% · 5 of 50 · the audited brand
  8. Five tied 0 to 18 or 20% · 3 of 50 each · Cherry, Das Keyboard, Ducky, Epomaker, ZSA

Eighth place is a five-way tie at 3 of 50: Cherry, Epomaker and ZSA run 0 to 18%, Das Keyboard and Ducky 0 to 20%; the bar shows the wider. Varmilo (2 of 50), Glorious and Razer (1 each) complete the 15 brands named.

The range covers both a cluster bootstrap over questions and a Wilson interval on the pooled count. With 10 questions it is a floor on the uncertainty; paid audits use 20 or more.

4 of the brand's 5 mentions came from one question, “What is the best low-profile mechanical keyboard?”, and it was never the first brand named. Asked the same question again, the engine named exactly the same tracked brands in 26 of 100 answer pairs. One answer is one draw, which is why every rate above carries a range.

Quoted answers

What the AI said

Three answers that named the brand, quoted exactly from the stored answers. The brand's name and product names are replaced by [the brand]; […] marks where a quote is cut.

  • Claude, web search

    What is the best low-profile mechanical keyboard?

    On a tighter budget, [the brand] or NuPhy Node75 Low-Profile get praised (including by RTINGS) for a surprisingly soft, cushioned typing feel and low input latency for the price. Bottom line: NuPhy for best all-around value, Corsair K100 AIR if money's no object, Logitech if gaming latency matters most.

    Question U01, answer 2 of 5, 2026-10-16

  • Claude, web search

    What is the best slim mechanical keyboard for programmers?

    If you want something more serious/customizable, the Keychron Q1 Max (low-profile variant) or [the brand] are worth a look […]

    Question U09, answer 1 of 5, 2026-10-16

  • Claude, web search

    What is the best low-profile mechanical keyboard?

    If gaming responsiveness matters most, the Corsair K100 AIR or Logitech G515 TKL are better suited. For tight budgets, [the brand] or NuPhy Node75 are solid.

    Question U01, answer 3 of 5, 2026-10-16

Sources

Cited sources and integrity flags

60 of 60 answers carried at least one cited URL, 4.0 per answer on average, pointing at 65 distinct domains. Of those 65 domains, 18 carry at least one flag (51 of 214 citation instances). Every flag is a prompt for a human look, not a finding about the site.

Narrow screen: scroll the table sideways for more columns.

Domains cited in 6 or more of the 60 answers · engine: Claude with web search · collected 2026-10-16
Domain Answers citing it Registered Tranco rank Flag
tomsguide.com182000-08-034,699No flag
rtings.com152011-11-2321,603No flag
unikeyboards.com112022-12-24515,408No flag
narcosis-the-game.com92026-04-16Not in top 1MFlagged registered 2025 or later
switchandclick.com82020-01-06Not in top 1MNo flag
techradar.com82003-12-062,339No flag
mkbguide.com72026-01-18Not in top 1MFlagged registered 2025 or later
techrepublic.com71998-11-205,377No flag
tomshardware.com71997-09-113,203No flag
analyticsinsight.net62016-05-0427,863No flag
fypower.org62003-07-18Not in top 1MFlagged old domain, WHOIS changed recently
  • Registered 2025 or later: check who runs it and whether its lists disclose paid placement.
  • Old domain, WHOIS changed recently: an older, low-traffic domain whose registration record changed in the last six months. Check whether the site changed owner and topic.

The full report lists all 18 flagged domains with the check each one asks for. Answers citing a domain count all 60 answers, including the 10 to the two questions that name the brand.

Graded fix list

Four fixes, graded by evidence

Do these three first: Fix 3, then Fix 1, then Fix 2. Each grade names the evidence behind the technique in general, not proof that it will move this brand's count; the re-measurement is what tests that.

Checked on the brand's site first

Read-only requests with no JavaScript run, 2026-10-16. Both checks passed, so neither is on the fix list.

  • Product pages read without JavaScript. The product page returned HTTP 200 and, with no script run, showed the brand name, the product name, a price and 3,369 words of body text.
  • AI search, user-fetch and Gemini crawlers allowed in robots.txt. All 9 are allowed for the home page and the product page, among them OAI-SearchBot, ChatGPT-User, Claude-SearchBot, PerplexityBot, Google-Extended and Bingbot. Training-only crawlers (GPTBot, ClaudeBot) are allowed too, which is a choice rather than a fix.
  1. Fix 1

    Add sourced figures and named quotes to your key pages

    Technique
    Statistics, quotations and citations embedded in the brand's own copy (the GEO paper tactics).
    Evidence grade
    Controlled study. KDD 2024 GEO paper: +28% to 41% visibility per intervention, up to 40% combined, 37% in a live Perplexity test.
    Cost
    Very low; 1 to 2 hours per page.
    Risk
    Low; accuracy and FTC exposure if a statistic is sponsored.
    What this audit shows
    The brand's own site was cited in 3 answers, none of them to a question that names no tracked brand. In this run the engine never cited it when answering a category question, so own-page edits are the slower route to these answers; the third-party pages in Fix 3 are the direct one.
    First step
    Pick the product page you most want recommended. Add one measured figure with its source linked (a spec from a named test, a battery-life measurement, a weight) and one quoted line from a named review with the publication credited, both in the first screen of text.
  2. Fix 2

    Refresh your key pages on a schedule

    Technique
    Content refresh on a cadence (substantive updates, not date changes).
    Evidence grade
    Large correlational study. Ahrefs, 17M AI citations: cited content runs 25.7% fresher than organic results; 76.4% of ChatGPT's top citations were under 30 days old. Corroborated by Seer Interactive.
    Cost
    Low; 1 to 3 hours per page.
    Risk
    Minimal; a cosmetic-only refresh is a negative signal.
    What this audit shows
    The brand's own site was cited in 3 answers, none of them to a question that names no tracked brand. In this run the engine never cited it when answering a category question, so own-page edits are the slower route to these answers; the third-party pages in Fix 3 are the direct one.
    First step
    List the 20 pages that matter most (best sellers and comparison pages). Refresh 10 of them now with a substantive change (a spec, a price, a new comparison, a new review quote) and leave the other 10 as they are, so the 90-day re-measurement can tell whether the refresh helped. Update the visible last-updated date only where the content changed.
  3. Fix 3

    Get onto the buying guides the AI already cites

    Technique
    Inclusion in existing third-party ‘best X’ lists and reviews (earned, not self-published).
    Evidence grade
    Large correlational study (three independent studies converge). Ahrefs, Wix Studio and Evertune: 21.9% to 63% of AI citations go to listicles; self-promotional listicles get 65% fewer citations.
    Cost
    Outreach time; free to about $2,000 per placement where paid (paid placement needs disclosure).
    Risk
    FTC disclosure if paid; a free sample is a material connection the reviewer must disclose; exposure if the host site is later penalised.
    What this audit shows
    The engine cited the pages below in answers that named a competitor but not the brand.
    First step

    Open each page below and check whether it already covers the brand; where it does not, send the pitch for that page from the outreach kit to the author it names, or to the site's contact page.

    Pages to pitch first: for each site, the page cited most often in answers that named a competitor but not the brand, and how many such answers cited the site.

    Not pitched with this note: rtings.com, tomsguide.com, techradar.com and the-gadgeteer.com already review the brand. rtings.com buys every product it tests, so a sample offer misfires there; ask to be considered for a future round of testing instead.

  4. Fix 4

    Check whether AI crawlers visit your site at all

    Technique
    Log-file analysis of AI crawler traffic.
    Evidence grade
    Confirmed mechanism (direct HTTP observation). Cloudflare and Vercel publish crawl-to-referral ratios: ClaudeBot 23,951 to 1, GPTBot 1,276 to 1, PerplexityBot 111 to 1.
    Cost
    Free to low; 1 to 2 hours to set up.
    Risk
    Blocking a training bot is not the same as blocking a citation bot (see the crawler check above).
    What this audit shows
    Nothing; this needs the brand's server logs, which an outside audit does not see.
    First step

    Get the last 30 days of web server access logs from your host (on a hosted store without raw logs, ask the host or use your CDN's bot report) and count the AI crawlers by name:

    grep -Eo "GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-SearchBot|Claude-User|PerplexityBot|Perplexity-User" access.log | sort | uniq -c

Competitor content gap

Who the cited guides name, and what the engines repeat

This section is included in the Category and Market Studies.

What the engines say about each brand, and what the pages they cite say. Phrases come from the sentence around each mention in the answers to the 10 questions that name no brand: product names, prices, specs, evaluative words and review sources, each counted once per answer and only when the engine wrote it. Pages are the 10 the engines cited most on those questions (tracked brands' own sites left out), read without running JavaScript.

Narrow screen: scroll the table sideways for more columns.

The audited brand and its 8 tracked competitors · guide pages read 2026-10-17 · answers: engine Claude with web search
Brand Named on the 10 most-cited guide pages Answers naming it (questions naming no brand) Three phrases repeated most in those answers
The brand (audited)3 of 9 read5[the brand's model] (3 answers); gasket-mounted (2); keycaps (2)
Keychron9 of 9 read45hot-swappable (28 answers); wireless (21); wired (17)
NuPhy4 of 9 read31low-profile (29 answers); NuPhy Air75 V2 (17); slim (14)
Logitech7 of 9 read26tactile (18 answers); low-profile (16); Logitech MX Mechanical (15)
Royal Kludge5 of 9 read14Royal Kludge RK87 (5 answers); hot-swap (5); wireless (5)
Akko0 of 9 read6Cherry MX (3 answers); silent (3); Akko FUN60 Ultra (2)
Epomaker3 of 9 read3Epomaker TH80 Pro (2 answers); keycaps (1); OLED (1)
ZSA1 of 9 read3ZSA Voyager (3 answers); split (3); ergonomic (2)
Kinesis1 of 9 read0None repeated

Not read, so counted as neither naming nor skipping a brand: techrepublic.com (cited in 7 answers; HTTP 403).

Narrow screen: scroll the table sideways for more columns.

The 10 pages, by answers citing them on questions that name no brand · first five of 10 shown
# Page Answers citing it Read Audited brand and competitors named, in page order Names the brand Source flag
1techrepublic.com, “Best Mechanical Keyboards of 2026: 5 Picks for Gaming…”7No (HTTP 403)n/an/aNo flag
2analyticsinsight.net, “Best Mechanical Keyboards for Coding and Writing in 2026”6YesKeychron, NuPhy, ZSA, LogitechNoNo flag
3narcosis-the-game.com, “13 Best Low-Profile Mechanical Keyboards (October 2026)…”6YesLogitech, NuPhy, the brand, Keychron, Royal KludgeYesFlagged registered 2025 or later
4rtings.com, “RTINGS: Best Low Profile Keyboards”6YesLogitech, Keychron, NuPhy, the brandYesNo flag
5igeeksblog.com, “Best mechanical keyboards for Mac in 2026”5YesKinesis, Logitech, Keychron, EpomakerNoNo flag

Keychron is named on all 9 of the most-cited guide pages that could be read, the brand on 3, and 2 of those 3 carry a source flag. The full section goes on to pages 6 to 10, the sentence where each page first names the brand, and the gap in six plain lines, among them: no answer to a category question quoted a price with the brand, while a price came with Keychron in 17.

Outreach kit

Who wrote the guides, and what to send them

This section is included in the Category and Market Studies.

For each page in Fix 3 (the pages the engine cited in answers that named a competitor but not the brand): who wrote it, where to reach them, and a pitch of at most 70 words. A name comes only from the page itself (schema.org author in its structured data, a meta author tag, a rel=author link or a visible “By” line); a page without one is sent to the site's contact page instead. Every pitch ends with the disclosure line, because a free unit is a material connection the reviewer must disclose.

Two of the four pages are shown. The report prints each author's name; it is withheld here.

  • Page 1

    unikeyboards.com

    Page
    The Best 60% Keyboards of 2026, cited in 5 answers that named a competitor but not the brand, 11 for the site as a whole; most often for U06, “Best compact mechanical keyboard to carry around with a laptop?”
    What it covers
    Read 2026-10-17 09:45 PDT, HTTP 200: titled “The Best 60% Keyboards of 2026”; names Ducky, Keychron and Razer; does not name the brand.
    Author
    Named on the page (withheld here), read from schema.org author in the page's JSON-LD.
    Site contact
    Site contact page, if the author has no public address: unikeyboards.com/policies/contact-information
    Pitch

    Hello [author], I run [the brand]. Your guide “The Best 60% Keyboards of 2026” covers Ducky, Keychron and Razer but not us. We make [the product], listed at [price] on [the brand's site]. I can send a review unit to test, with no conditions on what you write. A free unit is a material connection; please disclose it as your policy requires.

    Written in the brand's first person, because the brand sends it. The report's version is 63 words, with the name, product, price and site filled in.

  • Page 3

    tomshardware.com

    Page
    Best Budget Mechanical Keyboards 2026, cited in 4 answers that named a competitor but not the brand, 7 for the site as a whole; most often for U10, “What is a good mechanical keyboard to give as a gift, around $100 to $150?”
    What it covers
    Read 2026-10-17 09:45 PDT, HTTP 200 (the address now redirects to the site's budget mechanical keyboards page): titled “Best Budget Mechanical Keyboards 2026”; names Razer, Corsair, Logitech, Keychron, Royal Kludge, Epomaker and Glorious; does not name the brand.
    Author
    Named on the page (withheld here), read from schema.org author in the page's JSON-LD.
    Site contact
    Site contact page, if the author has no public address: tomshardware.com/reviews/about-us,4260.html
    Pitch
    In the report, 62 words in the same shape: the guide's title, the first three brands it names (Razer, Corsair and Logitech), the same offer and the same disclosure line.

The other two pages, switchandclick.com and analyticsinsight.net, get the same five lines. In this audit all four pages named an author and linked a contact page, and the four pitches run 62 to 66 words.

Page briefs

Five of the brand's pages, read the way a crawler reads them

This section is included in the Category and Market Studies.

The pages of the brand's own site the engines cited, most cited first, then the product or category pages its homepage links to most among those matching the audit's category, each read as a crawler that runs no JavaScript reads it, main content only. For each: what it already has, then five items in a fixed order, each marked already in place or add, with the evidence read.

Page 1

The product page the engines cite most

Product page, read 2026-10-17 10:16 PDT, HTTP 200, 3,210 words of main text; chosen as cited in 3 answers in this run.

What it already has: a visible price ([price]); structured data: AggregateRating, Brand, BreadcrumbList, ListItem, Offer, Person, Product, ProductGroup, Rating, Review; named specs: 96%, 100 keys, 40 gf, 2.4 GHz, 84 Keys, 1000 Hz; review quotes: none credited to a publication; plain-text questions: 0.

  1. Add 1. One sourced, measured figure

    Evidence: 12 figures stated (96%, 100 keys, 40 gf, 2.4 GHz), none with a named test or source beside it.

    To do: Add one measured figure for [the product] and say who measured it, with a link. The engines quote measured figures for rivals: NuPhy's weight (2.2 lb, 3 answers) and Keychron's battery life (200 hours, 2 answers).

  2. Add 2. One quoted line from a named review, publication credited

    Evidence: no quotation credited to a publication in the main text (customer reviews do not count).

    To do: Quote one line from microcenter.com (its page for [the product], cited in 1 answer in this audit) and credit microcenter.com by name.

  3. Add 3. A comparison table against the two most-named competitors

    Evidence: 0 tables in the main text; Keychron and NuPhy not named.

    To do: Add a table: [the product] against Keychron K8 Pro and NuPhy Air75 V2 (the models the engines name most for each), with rows for price and hot-swappable, wireless, wired and aluminum, the points the engines repeat about them. Fill each cell from a source you can name.

  4. Add 4. Three plain-text question-and-answer pairs from the audit's own questions

    Evidence: 0 plain-text questions in the main text.

    To do: Add these three from the audit, each as a heading with a two- or three-sentence answer in the page text (not only inside a script or a tab that needs JavaScript): U01 “What is the best low-profile mechanical keyboard?”; U10 “What is a good mechanical keyboard to give as a gift, around $100 to $150?”; U07 “Which mechanical keyboard brands make the best keyboards for writers?”. Answer from what the page already states ([price]).

  5. Already in place 5. Structured data, where missing

    Evidence: JSON-LD types read: AggregateRating, Brand, BreadcrumbList, ListItem, Offer, Person, Product, ProductGroup, Rating, Review; price in structured data: yes.

The other four pages, the category page and three further product pages, get the same five items. Across all five: 20 items to add, 4 already in place, 1 not applicable (structured data on the category page, which needs none of the markup the research covers).

90-day roadmap

Nine items, in the order to start them

This section is included in the Category and Market Studies.

Report date October 17, 2026; re-measure on or after January 15, 2027, 90 days later. Every fix above and every page-brief item marked add, in the order to start them. Hours come from the techniques research (its cost and time column); where it gives no range the words say so and the item stays out of the total. Who: the client, the client's agency, or us under guided implementation (the client's team does the work, we sequence and review).

Narrow screen: scroll the table sideways for more columns.

90-day roadmap · report date 2026-10-17 · re-measure on or after 2027-01-15
# Item Effort (hours) Who Expected effect (evidence grade) Re-measure
1Fix 3: get onto the buying guides the AI already cites (4 pitches in the outreach kit)5 to 10 a week while pitching (the research's figure for founder-led outreach, A1, the same channel; it gives no hours for A9 itself)ClientPuts the brand on the third-party lists the engines already cite for these questions; listicles took 21.9% to 63% of AI citations across three studies. Evidence: Large correlational study (three independent studies converge).2027-01-15
2Fix 1: add sourced figures and named quotes to your key pages (through the page briefs)Counted in the page-brief rows belowClientGives the engine sentences with numbers and named sources to lift; the GEO study measured 28% to 41% more visibility per tactic on its own test set, not on this brand. Evidence: Controlled study.2027-01-15
3Fix 2: refresh your key pages on a schedule (the five briefed pages can count toward the 10)1 to 3 a page; 10 pages: 10 to 30ClientKeeps key pages inside the recency window the engines favour (76.4% of ChatGPT's top citations were under 30 days old in the Ahrefs study). Evidence: Large correlational study.2027-01-15
4Fix 4: check whether AI crawlers visit your site at all1 to 2 to set upUs (guided implementation)Shows within 30 days whether AI crawlers reach the site at all; it does not move the count by itself. Evidence: Confirmed mechanism (direct HTTP observation).2027-01-15
5Page briefs, one sourced, measured figure (all 5 briefed pages)1 to 2 a page, together with the review quote; 5 pages: 5 to 10ClientGives the engine a checkable number with its source to lift; the GEO study measured 28% to 41% more visibility per tactic on its own test set. Evidence: Controlled study.2027-01-15
6Page briefs, one quoted line from a named review, publication credited (all 5 briefed pages)Inside the figure's 1 to 2 hoursClientQuotations were one of the GEO study's tactics (same study, same caveat). Evidence: Controlled study.2027-01-15
7Page briefs, a comparison table against the two most-named competitors (all 5 briefed pages)3 to 7 a page (1 to 3 of research, 2 to 4 of writing); 5 pages: 15 to 35Client's agencyGives a comparison question a page to cite; one vendor report puts 51% of non-homepage AI citations on comparison listicles, method not independently checked. Evidence: Practitioner case study.2027-01-15
8Page briefs, three plain-text question-and-answer pairs from the audit's own questions (all 5 briefed pages)0.5 to 1 a page; 5 pages: 2.5 to 5ClientA question heading with the answer in its first sentence is a passage an engine can lift whole; Google describes passage retrieval, the size of the effect is unmeasured. Evidence: Practitioner case study.2027-01-15
990-day re-measurementAbout 2 of ours (from our pricing sheet, not the techniques research)Us (guided implementation)Tells movement from noise: the same questions, engines, repeats and method as this audit. Evidence: Measurement, not a technique.2027-01-15

Total where the research gives hours: 33.5 to 82 hours over the 90 days, plus Fix 3 and the re-measurement, whose effort the table states in words.

Pre-registered expectation

Written before any change ships. What we expect to move: the brand's count on the same 10 questions that name no brand, today 5 of 50 answers (between 0 and 28 of every 100), and the number of the most-cited guide pages that name the brand, today 3 of 9 read. We expect the change first on U06, U07 and U10, the questions the pitched pages were cited for. What would count as no effect: on the re-measurement on or after January 15, 2027, the brand named in fewer than 10 of 50 answers (10 is the smallest count whose 95% Wilson range sits wholly above today's 10 per 100); for the guide pages, no page that skips the brand today naming it then. With this many questions and repeats only a large change clears that bar; a smaller one reads as noise, which is why the bar is set now and not after the numbers are in.

Limits

What we will not promise

  • A ranking, a position or a placement in any AI answer. The engines decide; this audit measures.
  • That any fix above moves the count. Each carries its evidence grade; the 90-day re-measurement on the same questions and method is what tells movement from noise.
  • A single score. Every number here is a range with its repeat count, because one answer is one draw.
  • Coverage of engines or tiers that were not sampled. The sampling record lists exactly what was.
  • Any of the following, which this practice declines outright: bought or incentivised reviews; undisclosed Wikipedia edits; incentivised or sock-puppet Reddit posts; site networks and paid link schemes; hidden text or cloaked pages served to AI crawlers; guaranteed placement in any AI answer.

What it costs

This is the shape of a Category Study.

This demonstration is a small sample run on one assistant. The Category Study runs 2,400+ simulated buyer conversations across 60 buyer scenarios, in two waves. The Brand Study stops at the graded fix list; the four sections after it come with the Category and Market Studies, and the Market Study adds a test plan written before any change ships.

Compare the three studies

Get the free snapshot