the academic backbone · ai tracker

Methodology & Validation

how every number is measured, what changed and when, and how our numbers compare to vendor ground truth — the audit trail behind the dataset.

updated 2026-08-01 · search-share data through Jul 19, 2026 · refreshed daily · this page is fully baked — no JavaScript needed

01 · the exact measurement

tool registry, payload groups & anchor normalization

Google Trends compares at most 5 terms per payload, so the 26 tools are pulled in 7 payload groups (A–G) with Gemini AI as the shared anchor term in every group. share of Google search interest — relative preference, not usage counts.

tool	company	exact search term	payload group	id
ChatGPT	OpenAI	`ChatGPT`	A	`chatgpt`
Gemini	Google	`Gemini AI`	A · anchor in all 7 groups	`gemini`
Claude	Anthropic	`Claude AI`	B	`claude`
DeepSeek	DeepSeek	`DeepSeek`	A	`deepseek`
Grok	xAI	`Grok`	A	`grok`
Meta AI	Meta	`Meta AI`	A	`meta_ai`
Perplexity	Perplexity	`Perplexity AI`	B	`perplexity`
Microsoft Copilot	Microsoft	`Microsoft Copilot`	B	`ms_copilot`
Midjourney	Midjourney	`Midjourney`	B	`midjourney`
GitHub Copilot	GitHub	`GitHub Copilot`	C	`gh_copilot`
Cursor	Anysphere	`Cursor AI`	C	`cursor`
Character.AI	Character	`Character AI`	C	`character_ai`
NotebookLM	Google	`NotebookLM`	C	`notebooklm`
Janitor AI	JanitorAI	`Janitor AI`	D	`janitor_ai`
Leonardo.Ai	Canva	`Leonardo AI`	D	`leonardo`
Suno	Suno	`Suno AI`	D	`suno`
Runway	Runway	`Runway AI`	D	`runway`
ElevenLabs	ElevenLabs	`ElevenLabs`	E	`elevenlabs`
Sora	OpenAI	`Sora AI`	E	`sora`
Kimi	Moonshot AI	`Kimi AI`	E	`kimi`
Qwen	Alibaba	`Qwen`	E	`qwen`
Mistral	Mistral AI	`Mistral AI`	F	`mistral`
Poe	Quora	`Poe AI`	F	`poe`
Luma	Luma AI	`Luma AI`	F	`luma`
Replit	Replit	`Replit`	G	`replit`
Veo	Google	`Google Veo`	G	`veo`

anchor normalization — written out

for each payload group G in {B, C, D, E, F, G}, per country r:

  scale_G(r) = A[r]["Gemini AI"] / G[r]["Gemini AI"]     when both anchor readings ≥ 2
             = mean of that ratio over all countries          otherwise (group-mean ratio;
               where both anchors ≥ 2                          1.0 if no such country)

  value[r][tool] = raw_G[r][term] × scale_G(r)      (duplicate anchor columns dropped;
                                                     gemini's value comes from group A only)

  share[r][tool] = value[r][tool] / Σ value[r][·] × 100     → the 26 tools sum to ~100% per country

time series: identical logic PER TIMEPOINT (anchor ≥ 2 test, else the series-wide mean ratio).

timeframes & cadence

Country map: now 7-d — interest_by_region, COUNTRY resolution, inc_low_vol=False; a rolling 7-day window snapshotted daily.
Global race, weekly: today 12-m — interest_over_time, one point per week over 12 months.
Global race, daily: today 3-m — interest_over_time, one point per day over 3 months.
Cadence: one snapshot per day (~06:30 site-local time), filenames dated by UTC day; Trends pulls older than 72 h are treated as missing, never reused.

02 · the five signal families

every endpoint, exactly as queried

all five families are fetched by the same daily job and archived as raw snapshots under /history/ — see the research console for the full data dictionary and API.

family	source & exact endpoint	cadence
search preference	Google Trends via pytrends — interest_by_region (resolution=COUNTRY, inc_low_vol=False, timeframe “now 7-d”) for the country map; interest_over_time “today 12-m” (weekly) and “today 3-m” (daily) for the global race; 7 payload groups of ≤5 terms, “Gemini AI” anchored in each	daily
SDK downloads	npm registry — https://api.npmjs.org/downloads/point/last-week/{pkg} and https://api.npmjs.org/downloads/range/{start}:{end}/{pkg} (8 packages); PyPI — https://pypistats.org/api/packages/{pkg}/recent and …/overall?mirrors=false (11 packages)	daily
OSS momentum	GitHub REST — https://api.github.com/repos/{owner}/{repo}, stargazers_count (17 flagship AI repos)	daily
open models	Hugging Face — https://huggingface.co/api/models?sort=downloads&direction=-1&limit=20 and …?sort=trendingScore… (top-downloaded + trending snapshots)	daily
attention	Wikimedia REST — https://wikimedia.org/api/rest_v1/metrics/pageviews/per-article/en.wikipedia/all-access/user/{article}/daily/{start}/{end} (user, non-bot views; 365-day backfill; article titles redirect-resolved via https://en.wikipedia.org/w/api.php)	daily

03 · known limitations

what this data cannot tell you

the caveats that belong in any methods section built on this dataset — stated here once, plainly.

Trends sampling & rescaling opacity. Google Trends values are a sampled, 0–100-quantized estimate; the same query can return slightly different numbers on different days and Google does not document the sampling. The shared anchor and daily pulls dampen this, but day-to-day jitter of a point or two is expected and real.
Relative, not absolute. A 40% share says nothing about how many people searched — only what a country's AI-engaged searchers preferred among the 26 tracked tools. Search interest ≠ usage.
1-week windows. Country shares come from a single rolling week (“now 7-d”), so news-cycle spikes move them; the Anthropic AEI benchmark used in validation is likewise built from ~1-week Claude.ai usage windows. Both are snapshots, not long-run averages.
Low-volume countries excluded. inc_low_vol=False means only ~85 countries clear Trends' volume floor; countries absent from the map are unmeasured, not zero.
Anchor threshold. Cross-group scaling trusts the per-country anchor only when both readings are ≥ 2; below that a group-mean ratio substitutes, so group-B/C values in small countries are noisier than group-A values.
CI-inflated downloads. npm and PyPI weekly downloads count CI machines, mirrors and bots — momentum, not headcount. PyPI is pulled with mirrors=false, which removes only known mirrors.
Redirect-era Wikipedia attribution. Wikimedia attributes pageviews before an article rename to the old title, so early segments of renamed articles (e.g. “Claude (AI)” before ~mid-2026) understate attention. Titles are redirect-resolved at fetch time since v1.1.
Term ambiguity. Search terms like “Grok” or “Cursor AI” can capture unrelated interest; terms are frozen in the registry above and only change with a changelog entry.
Stars are cumulative. GitHub star counts are lifetime totals — day-over-day deltas are the momentum signal; stars ≠ users.
The archive starts 2026-07-07. Country-level daily history begins at the archive start and is not retroactively reconstructible from Google Trends.

04 · validation study

construct validity — how our numbers compare to vendor ground truth

Spearman rank correlation (primary; robust, valid for ordinal data) and Pearson on log-transformed values where both measures are continuous. Country names/ISO codes joined via explicit alias table. Pure-python; script: scripts/validation_study.py. Stanford HAI 2026 excluded: its country genAI figures republish Microsoft's (not independent).

Our Claude preference share vs Anthropic's AI Usage Index

ourswithin-country Claude share of AI-tool search interest (%) theirsAUI recomputed from Anthropic's Feb-2026 usage shares — Claude.ai usage per working-age capita, world-proportional = 1.0

full samplen = 62 · Spearman ρ = −0.164 · Pearson (log) = −0.082

high-adoption countries (AUI ≥ 1.0)n = 25 · Spearman ρ = 0.587 · Pearson (log) = 0.585

source: Anthropic Economic Index · publisher: Anthropic · license: MIT · accessed 2026-07-08 · 62 countries plotted

honest readFull-sample agreement is weak BY CONSTRUCTION: our share measures preference among a country's AI-engaged searchers, Anthropic's index measures population-wide usage intensity. In low-adoption countries the engaged minority skews developer-heavy — and Claude is developer-favored — so Claude's preference share runs high exactly where per-capita usage is low. Where adoption is substantial (AUI ≥ 1), the two measures agree well; we report both rather than hiding either.

Our ChatGPT preference share vs OpenAI's per-capita usage rank

ourswithin-country ChatGPT share of AI-tool search interest, as rank theirsOpenAI Signals: per-capita ChatGPT message-usage rank (1 = highest), 2026-01 quarter

full samplen = 77 · Spearman ρ = 0.458

source: OpenAI Signals · publisher: OpenAI · license: CC BY 4.0 · accessed 2026-07-08 · cite: NBER Working Paper 34255 · 77 countries plotted

honest readWeak construct match, reported for completeness: OpenAI ranks countries by per-capita usage INTENSITY; our share measures ChatGPT's WITHIN-COUNTRY dominance over other AI tools. A country can use little AI overall yet prefer ChatGPT overwhelmingly, so low correlation here does not indicate error in either dataset.

Our developer density vs Microsoft's AI Diffusion rate

oursmap signal: GitHub developers per 1,000 working-age people (population data: Anthropic AEI) theirsMicrosoft AI Diffusion — % of working-age population using generative AI (Q1 2026)

full samplen = 124 · Spearman ρ = 0.723 · Pearson (log) = 0.713

source: Microsoft AI Diffusion Report · publisher: Microsoft AI Economy Institute · license: MIT · accessed 2026-07-08 · 124 countries plotted

honest readOur map counts GitHub-registered developers (a developer-population proxy); Microsoft estimates general-population genAI use. Related but distinct: developer density should correlate with, not equal, population-wide adoption.

excluded on purpose

excluding non-independent data is part of the method — a benchmark that republishes another benchmark would only double-count agreement.

Stanford HAI AI Index 2026 — excluded: country-level genAI adoption figures are identical to Microsoft AI Diffusion H2-2025 — republished, not independent.

validation computed 2026-07-08 · the validation script (validation_study.py) and the full data dictionary ship inside the Zenodo deposit README, so every figure above is reproducible from the archive.

05 · methodology changelog

every change, on the record

this changelog is append-only: every change that touches how a number is computed gets an entry before it ships, and any change that breaks comparability of a series with its own past is flagged as a series break with its exact scope. comparing values across time? check the breaks first. newest first.

v3.1 Tracker surface expanded: head-to-head + per-country pages 2026-07-19

66 head-to-head pages added (/tracker/<a>-vs-<b>, ids alphabetical): one for every pair among the top 12 tools by current global search share, with baked verdict, who-prefers-what country tables and computed FAQ. The pair set follows the top-12 ranking, so pages appear/retire as the ranking moves; retired pages are pruned from the site and sitemap.
83 country pages added (/tracker/country/<iso2>): every tracked country's full ranked tool table, stat tiles, per-country timeline and computed FAQ.
Per-tool pages now generate for all 27 published tools straight from the content/tracker-tools.json registry (previously a hand-maintained 13); the today board, developer-signal panels and methodology spec table read the same registry.
Derived views over the same snapshots and shares — no metric definition changed and no new measurement was introduced.

v3.0 Tracked tool set expanded 13 → 27 (single-source registry) 2026-07-19 series break

series-break scope: ALL search-share series (per-country and global weekly/daily): denominator re-based from 13 to 27 tools on 2026-07-19

14 tools added: Janitor AI, Leonardo.Ai, Suno, Runway, ElevenLabs, Sora, Kimi, Qwen, Mistral, Poe, Luma, Windsurf, Replit, Veo. The tracked set now lives in content/tracker-tools.json (ids, search terms, payload groups, wiki articles, official packages/repos) — the single source of truth read by fetch_tools.py, fetch_signals.py and build_tracker.py.
Search preference now pulls 7 payload groups (A–G, ≤5 terms each; unchanged groups A/B/C keep their exact terms and order) with the same shared anchor term 'Gemini AI' first in every new group and identical anchor cross-normalization.
SERIES BREAK (all search-share series, country and global): shares are percentages of the tracked-set total, and the denominator grew from 13 to 27 tools on 2026-07-19 — every share steps DOWN on this date (e.g. US ChatGPT 78.9 → 75.6). To compare across the break, renormalize post-break shares to the original 13-tool subset (all 13 ids remain published every day).
Startup grace rule added: a NEWLY added tool whose share rounds to 0.0 in every country is excluded from that day's output (logged loudly) instead of failing the publish gate; the original 13 tools are never dropped. The published set is recorded per day in the data ('tool_ids'). On 2026-07-19 all 27 tools had signal and none were dropped.
Attention: 11 new Wikipedia articles begin with the standard 365-day backfill (no break for existing articles; tools without a dedicated article — Janitor AI, Poe, Windsurf — carry no wiki series). Robustness fix: the Wikimedia pageviews API rejects a date range ending past an article's last loaded day (per-article load lag of 1–3 days is routine), so the end date now walks back up to 3 days on 404 — articles no longer flap out of the dataset for a day (previously affected e.g. 'Meta AI'). No effect on stored values.
Developer signal (additive, no break): npm +@elevenlabs/elevenlabs-js; PyPI +elevenlabs; GitHub +suno-ai/bark, +elevenlabs/elevenlabs-python, +MoonshotAI/Kimi-K2, +QwenLM/Qwen3, +mistralai/mistral-inference.

v2.0 Multi-page tracker + research bundles 2026-07-08

Tracker restructured from a single page into /tracker hub + 13 per-tool pages + /tracker/compare + /tracker/developers; no metric definitions changed.
Per-country rollup series (history/series/<ISO2>.json) and yearly research bundles (history/research/) introduced — derived views over the same snapshots, no new measurement.
Research console (/tracker/research) added: filterable slices, CSV/JSON export with embedded caveats, reproducible query permalinks, documented free JSON API.

v1.1 Data-quality fixes from pre-launch adversarial review 2026-07-08 series break

series-break scope: wiki attention for gemini, claude, cursor (backfill re-based to canonical articles)

Wikipedia article titles now auto-resolve redirects to canonical titles before fetching pageviews. SERIES BREAK (wiki attention, tools: gemini, claude, cursor): values fetched before this date used redirect titles ('Gemini (chatbot)', 'Claude (language model)', 'Cursor (code editor)') whose pageviews were near zero; the backfilled series now reflect the canonical articles ('Google Gemini', 'Claude (AI)', 'Cursor (company)'). Note: Wikimedia attributes pre-rename pageviews to the old titles, so early segments of a renamed article's series understate attention (e.g. 'Claude (AI)' before ~mid-2026).
npm daily download series now end at the last COMPLETE UTC day; the npm range API reports 0 for the in-progress day and those artifact zeros previously appeared as a same-day collapse. No effect on true values.
Snapshot filenames switched from local to UTC dating (prevents a same-data snapshot being written under two different dates around midnight).
Stale-cache caps added: Trends pulls older than 72h and signal-family caches older than 7 days are no longer reused; carried-over halves no longer restamp the artifact 'updated' time.
Taiwan ISO2 join fixed on the map (display only; data unchanged).

v1.0 Initial methodology — daily archive begins 2026-07-07

13 tools tracked: ChatGPT, Gemini, Claude, DeepSeek, Grok, Meta AI, Perplexity, Microsoft Copilot, Midjourney, GitHub Copilot, Cursor, Character.AI, NotebookLM.
Search preference: Google Trends interest_by_region (timeframe 'now 7-d', inc_low_vol=False), pulled in three payload groups of ≤5 terms with 'Gemini AI' as the shared anchor term in every group; group values cross-normalized via the per-country anchor ratio (fallback: group-wide mean ratio when a country's anchor value < 2), then converted to per-country shares summing to ~100.
Global race series: interest_over_time 'today 12-m' (weekly) and 'today 3-m' (daily), anchor-normalized per timepoint, converted to shares.
Developer signal: npm registry weekly downloads (8 packages) + PyPI weekly downloads via pypistats (10 packages) + GitHub star counts (12 flagship AI repos) + Hugging Face top-downloaded and trending model snapshots.
Attention: Wikipedia daily user pageviews per tool article, 365-day backfill.
Daily country-level snapshots begin 2026-07-07 (not retroactively reconstructible from Google Trends).

06 · how to cite

put it in the bibliography

DOI: 10.5281/zenodo.21252715 (concept DOI 10.5281/zenodo.21252714).

APA

PIXIPACE. (2026). Who is using AI? — Global AI Tracker dataset [Data set]. whoisusingai.com. Retrieved Aug 1, 2026, from https://whoisusingai.com/tracker/research https://doi.org/10.5281/zenodo.21252715

BibTeX

@misc{whoisusingai2026tracker,
  author       = {{PIXIPACE}},
  title        = {Who is using AI? --- Global AI Tracker dataset},
  year         = {2026},
  url          = {https://whoisusingai.com/tracker/research},
  doi          = {10.5281/zenodo.21252715},
  note         = {Accessed 2026-08-01}
}

the data and the plain-JSON API live on the research console. the collection and build scripts are not published on this site — the validation script and the full data dictionary ship inside the Zenodo deposit README instead. search shares are relative Google search interest, not usage counts — say so in your methods section.