"""Answer the README's question from the fetched JSON. Writes results.md and monthly.csv. python research/wikipedia-traffic/fetch.py # once python research/wikipedia-traffic/analyse.py Every number in results.md and in the journal entry comes from here. """ import csv import json import os import statistics import sys RAW = os.path.expanduser("~/data/wikipedia-traffic") HERE = os.path.dirname(os.path.abspath(__file__)) BN = 1e9 M = 1e6 def series(project, access, agent): """{'YYYY-MM': views} from one fetched file.""" path = os.path.join(RAW, "%s__%s__%s.json" % (project, access, agent)) with open(path) as fh: data = json.load(fh) out = {} for item in data["items"]: stamp = item["timestamp"] # YYYYMMDDHH key = "%s-%s" % (stamp[:4], stamp[4:6]) out[key] = int(item.get("views", item.get("devices", 0))) return out def devices(site): path = os.path.join(RAW, "unique-devices__%s.json" % site) with open(path) as fh: data = json.load(fh) return {"%s-%s" % (i["timestamp"][:4], i["timestamp"][4:6]): int(i["devices"]) for i in data["items"]} def prev(month, years=1): return "%d-%s" % (int(month[:4]) - years, month[5:]) def yoy(s, month): p = prev(month) if p not in s or month not in s or not s[p]: return None return s[month] / s[p] - 1.0 def w(s, months): return sum(s.get(m, 0) for m in months) def change(s, t12, p12): a, b = w(s, t12), w(s, p12) return (a / b - 1.0) if b else None def tail_negative_run(s, floor="2016-07"): """The unbroken run of negative year-over-year months ending at the last month.""" run = [] for m in reversed([m for m in sorted(s) if m >= floor]): v = yoy(s, m) if v is None or v >= 0: break run.append(m) return list(reversed(run)) def country_noise(): """Median absolute year-over-year change per country, from country.py's table. Empty when country.csv has not been generated. Section 8 is produced here rather than appended by hand, because this file writes results.md from scratch: anything added to that file by hand is destroyed by the next run. """ path = os.path.join(HERE, "country.csv") if not os.path.exists(path): return {} with open(path) as fh: rows = list(csv.DictReader(fh)) by_month = {r["month"]: r for r in rows} out = {} for country in [c for c in rows[0] if c != "month"]: changes = [] for month in by_month: a = by_month.get(prev(month), {}).get(country) b = by_month[month].get(country) if a and b and float(a): changes.append(abs(float(b) / float(a) - 1.0)) if changes: out[country] = statistics.median(changes) return out def main(): user = series("en.wikipedia", "all-access", "user") auto = series("en.wikipedia", "all-access", "automated") spider = series("en.wikipedia", "all-access", "spider") allproj = series("all-projects", "all-access", "user") by_access = {} for access in ("desktop", "mobile-web", "mobile-app"): by_access[access] = {g: series("en.wikipedia", access, g) for g in ("user", "automated", "spider") if os.path.exists(os.path.join( RAW, "en.wikipedia__%s__%s.json" % (access, g)))} months = [m for m in sorted(user) if m >= "2016-01"] latest = months[-1] nonhuman = {m: spider.get(m, 0) + auto.get(m, 0) for m in sorted(spider)} total = {m: user[m] + nonhuman.get(m, 0) for m in months} t12 = months[-12:] p12 = [prev(m) for m in t12] t12_change = change(user, t12, p12) # --- the pre-registered reclassification detector ----------------------- # Crude by construction: it compares a month-on-month fall in `user` with a # month-on-month rise in the non-human total within +/-1 month. Reported as # fixed, then tested a second way below, because it fires on seasonality. signature = [] ordered = sorted(user) for m in t12: i = ordered.index(m) drop = user[ordered[i - 1]] - user[m] if i else 0 if drop <= 0: continue for j in (i - 1, i, i + 1): if j <= 0 or j >= len(ordered): continue a, b = ordered[j - 1], ordered[j] if a in nonhuman and b in nonhuman and nonhuman[b] - nonhuman[a] >= 0.5 * drop: signature.append((m, drop, b, nonhuman[b] - nonhuman[a])) break if t12_change <= -0.10 and not signature: verdict = "ALIVE" elif t12_change > -0.03: verdict = "DEAD" else: verdict = "PARTLY" # --- monthly.csv -------------------------------------------------------- with open(os.path.join(HERE, "monthly.csv"), "w", newline="") as fh: wr = csv.writer(fh) wr.writerow(["month", "user", "automated", "spider", "nonhuman_share", "user_yoy", "desktop_user", "mobile_web_user", "mobile_app_user", "all_projects_user"]) for m in months: c = yoy(user, m) wr.writerow([m, user[m], auto.get(m, ""), spider.get(m, ""), "%.4f" % (nonhuman.get(m, 0) / total[m]) if total[m] else "", "%.4f" % c if c is not None else "", by_access["desktop"]["user"].get(m, ""), by_access["mobile-web"]["user"].get(m, ""), by_access["mobile-app"]["user"].get(m, ""), allproj.get(m, "")]) L = [] add = L.append add("# Results — is AI killing Wikipedia's traffic?") add("") add("Generated by `analyse.py` from Wikimedia's public REST pageview API. " "Complete calendar months; the last is **%s**. The question and the kill " "rule are in README.md and were fixed before the data was fetched." % latest) add("") add("## 1. The kill rule's answer") add("") add("| | human pageviews |") add("|---|---|") add("| trailing 12 months (%s..%s) | **%.2f bn** |" % (t12[0], t12[-1], w(user, t12) / BN)) add("| the same 12 months a year earlier | %.2f bn |" % (w(user, p12) / BN)) add("| change | **%+.1f %%** |" % (100 * t12_change)) add("") add("The pre-registered reclassification detector fires on **%d month(s)** of " "the window:" % len(signature)) for m, drop, b, rise in signature: add(" - `user` fell %.0f m into %s, non-human rose %.0f m into %s" % (drop / M, m, rise / M, b)) add("") add("**Verdict by the rule fixed before the data: %s** — the fall is inside " "the −3 %% to −10 %% band. Section 2 decides which kind of \"partly\" it is." % verdict) add("") add("## 2. Is it a counting change?") add("") add("The detector above compares month-on-month moves and so fires on ordinary " "seasonality. Two harder tests:") add("") nh_change = change(nonhuman, t12, p12) tot_change = change(total, t12, p12) add("**(a) If humans were being reclassified as bots, the non-human series would " "rise.** It falls:") add("") add("| series | trailing 12 (bn) | previous 12 (bn) | change |") add("|---|---|---|---|") for name, s in (("human (`user`)", user), ("non-human (`automated`+`spider`)", nonhuman), ("everything counted", total)): add("| %s | %.2f | %.2f | %+.1f %% |" % (name, w(s, t12) / BN, w(s, p12) / BN, 100 * change(s, t12, p12))) add("") add("Total counted traffic fell **%+.1f %%**, more than the human part. Traffic " "did not move between the classes; there is less of both." % (100 * tot_change)) add("") add("**(b) The largest discontinuity in the window is %s→%s, when the non-human " "count halved (%.2f bn → %.2f bn). Whatever caused it — the classifier changing, " "a bot campaign ending — if it had moved requests into or out of the human " "class, the human series would step at the same month.** It does not — " "September to October is a seasonal rise every year, and 2025's is inside the " "usual range:" % ("2025-09", "2025-10", nonhuman["2025-09"] / BN, nonhuman["2025-10"] / BN)) add("") add("| year | Sep → Oct, human pageviews |") add("|---|---|") for y in range(2016, int(latest[:4])): a, b = user.get("%d-09" % y), user.get("%d-10" % y) if a and b: add("| %d | %+.1f %% |" % (y, 100 * (b / a - 1))) add("") add("So: the fall in human pageviews is not an artefact of how Wikimedia sorts " "requests. It is fewer requests.") add("") add("## 3. Where the fall is: one access method, not all of them") add("") add("| access method | human, trailing 12 (bn) | previous 12 (bn) | change |") add("|---|---|---|---|") for access in ("desktop", "mobile-web", "mobile-app"): s = by_access[access]["user"] add("| %s | %.2f | %.2f | **%+.1f %%** |" % (access, w(s, t12) / BN, w(s, p12) / BN, 100 * change(s, t12, p12))) add("") add("Desktop reading of en.wikipedia by humans **grew**. The whole of the decline " "is the mobile web, plus a smaller fall in the app.") add("") add("The same split for the bots, which rules out a mobile-only classifier change " "as the explanation — mobile bot traffic falls by about as much as mobile " "human traffic, so nothing was moved between the two:") add("") add("| access method | non-human, trailing 12 (bn) | previous 12 (bn) | change |") add("|---|---|---|---|") for access in ("desktop", "mobile-web"): if "automated" not in by_access[access]: continue s = {m: by_access[access]["automated"].get(m, 0) + by_access[access]["spider"].get(m, 0) for m in sorted(by_access[access]["spider"])} add("| %s | %.2f | %.2f | %+.1f %% |" % (access, w(s, t12) / BN, w(s, p12) / BN, 100 * change(s, t12, p12))) add("") dk_h = w(by_access["desktop"]["user"], t12) dk_b = w({m: by_access["desktop"]["automated"].get(m, 0) + by_access["desktop"]["spider"].get(m, 0) for m in months}, t12) add("Worth stopping on: on **desktop**, %.2f bn of the last twelve months' " "pageviews were declared non-human against %.2f bn human — **%.1f %% of what " "en.wikipedia served to desktop URLs was not a person.** On the mobile web it " "is %.1f %%." % ( dk_b / BN, dk_h / BN, 100 * dk_b / (dk_b + dk_h), 100 * w({m: by_access["mobile-web"]["automated"].get(m, 0) + by_access["mobile-web"]["spider"].get(m, 0) for m in months}, t12) / (w({m: by_access["mobile-web"]["automated"].get(m, 0) + by_access["mobile-web"]["spider"].get(m, 0) for m in months}, t12) + w(by_access["mobile-web"]["user"], t12)))) add("") add("## 4. When it started") add("") add("| series | unbroken run of negative year-over-year months ending %s | starts |" % latest) add("|---|---|---|") for name, s in (("all human pageviews", user), ("mobile web, human", by_access["mobile-web"]["user"]), ("desktop, human", by_access["desktop"]["user"]), ("mobile app, human", by_access["mobile-app"]["user"])): run = tail_negative_run(s) add("| %s | %d months | %s |" % (name, len(run), run[0] if run else "—")) add("") mw = by_access["mobile-web"]["user"] add("The break is one month wide. Mobile web, human pageviews, year over year:") add("") add("| month | mobile web | desktop |") add("|---|---|---|") for m in ["2024-02", "2024-03", "2024-04", "2024-05", "2024-06", "2024-07"]: add("| %s | %+.1f %% | %+.1f %% |" % (m, 100 * yoy(mw, m), 100 * yoy(by_access["desktop"]["user"], m))) add("") add("April 2024 was the last month mobile web grew. It has not grown since. " "Desktop shows no such break.") add("") add("Google put AI Overviews in front of every searcher in the United States in " "May 2024. That is an alignment of one series with one date, and this data " "cannot turn it into a cause: the pageview API has no country split, the launch " "was American and en.wikipedia is read everywhere, and other things happened " "that spring. What the data does say is that whatever changed, changed for " "phones and not for desktops, in a single month, and has not reversed in %d months." % len(tail_negative_run(mw))) add("") add("## 5. The cross-check that disagrees, and why I trust it less") add("") add("Wikimedia also publishes **unique devices**, which is not a count of requests " "but an estimate from a first-party cookie:") add("") add("| series | trailing 12 (bn) | previous 12 (bn) | change |") add("|---|---|---|---|") dev_rows = {} for site in ("all-sites", "desktop-site", "mobile-site"): try: s = devices(site) except FileNotFoundError: continue dev_rows[site] = s add("| %s | %.2f | %.2f | %+.1f %% |" % (site, w(s, t12) / BN, w(s, p12) / BN, 100 * change(s, t12, p12))) add("") if "desktop-site" in dev_rows: add("Desktop **devices** are up %+.1f %% while desktop **pageviews** are up only " "%+.1f %%, which would mean every desktop reader suddenly read a fifth fewer " "pages. A cookie-based estimate moves when browsers change how they keep " "cookies, and that is the likelier reading. Where a server-side request count " "and a cookie estimate disagree, the request count is the harder number, and " "it is the one section 1 uses." % ( 100 * change(dev_rows["desktop-site"], t12, p12), 100 * change(by_access["desktop"]["user"], t12, p12))) add("") add("## 6. Context: the longer series") add("") add("| year | human pageviews (bn) | change |") add("|---|---|---|") prev_total = None for y in range(2016, int(latest[:4]) + 1): ms = [m for m in months if m.startswith(str(y))] tot = w(user, ms) if len(ms) == 12: c = "" if prev_total is None else "%+.1f %%" % (100 * (tot / prev_total - 1)) add("| %d | %.2f | %s |" % (y, tot / BN, c)) prev_total = tot else: base = w(user, [prev(m) for m in ms]) add("| %d (%s..%s) | %.2f | %+.1f %% vs the same months a year earlier |" % (y, ms[0], ms[-1], tot / BN, 100 * (tot / base - 1))) add("") add("The series has fallen before — 2021 and 2022 both lost ground and then 2023 " "took it back. What is new is the length: %d consecutive negative months." % len(tail_negative_run(user))) add("") add("All Wikimedia projects together, human pageviews: %.2f bn in the trailing " "twelve months against %.2f bn, **%+.1f %%**." % ( w(allproj, t12) / BN, w(allproj, p12) / BN, 100 * change(allproj, t12, p12))) add("") add("## 7. What this does not say") add("") add("Where the readers went. Nothing in a pageview count names a destination, and " "no number here should be read as \"they asked a chatbot instead\". The findings " "are: the fall is real and not a counting artefact; it is a phone fall; it began " "in May 2024 and has not stopped.") add("") noise = country_noise() if noise: add("## 8. The country proxy, and why it fails") add("") add("Section 4 says this API has no country split. That is loose: " "`top-per-country` exists, returning the fifty most-read articles in a " "country on a day with `views_ceil` rounded up. `country.py` sums the " "en.wikipedia articles in that top fifty, the 15th of each month, " "2022-01..2026-08, for eight countries (448 requests, cached).") add("") add("Median absolute year-over-year change per country:") add("") add("| " + " | ".join(noise) + " |") add("|" + "---|" * len(noise)) add("| " + " | ".join("%.1f %%" % (100 * v) for v in noise.values()) + " |") add("") add("The effect sought is a one-off step of about 11 %%. The noise floor " "equals it in the quietest country (%s, %.1f %%) and more than triples " "it in the loudest (%s, %.1f %%); single months exceed 100 %% in every " "country. US 2024-05 reads +3 %%, the third smallest absolute move of " "its 44 months — **not** evidence against an American cause, because a " "series this noisy is not evidence of anything." % (min(noise, key=noise.get), 100 * min(noise.values()), max(noise, key=noise.get), 100 * max(noise.values()))) add("") add("**Verdict: the proxy cannot resolve the question.** The limit stated in " "section 4 stands, now measured rather than asserted. Kept in the pack " "so nobody repeats it.") add("") text = "\n".join(L) + "\n" with open(os.path.join(HERE, "results.md"), "w") as fh: fh.write(text) print(text) return 0 if __name__ == "__main__": sys.exit(main())