Generative Engine Optimization · independent audit

binance.com in AI answers

In this probe, Binance was cited only when the question contained the word Binance. Across the unbranded crypto questions tested, the sources AI assistants return are Coinbase Learn, Kraken Learn and Crypto.com Learn. Binance Academy, the largest crypto education corpus in the industry, did not appear once in this probe.

Owned-citation rate
1/7
only the branded query
Unbranded queries won
0
of 5 tested
llms.txt
none
4 of 5 rivals ship one
AI crawlers named in robots
0
6 classic crawlers are

Independent analysis built from public data for a job application. Not affiliated with Binance. Every finding below is reproducible with the live tool.

00

Summary

Four findings, each verifiable in under a minute.

Finding 1 · demand side

The citation footprint is branded-only

Across 7 probe queries, Binance-owned pages were cited once, on is Binance safe. On all five unbranded questions tested (staking, funding rates, best exchange, buying BTC in Malaysia) the cited sources were competitors and affiliates. At n=7 this is a direction to investigate at scale, and the direction is consistent.

Finding 2 · supply side

No llms.txt, while rivals shipped theirs

Coinbase, OKX, Bybit and Crypto.com all serve a real, parseable llms.txt. OKX's runs 15 sections and 73 curated links. Binance serves nothing at that path from any vantage point I tested, and binance.info, which is not WAF-challenged, returns a normal 404 page for it. Kraken serves an HTML 404.

Finding 3 · the mechanical one

Crawl hygiene that AI crawlers never read

robots.txt groups are most-specific-wins and do not inherit. Binance wrote 68 disallow rules across groups naming Googlebot, Bingbot, Baiduspider, DuckDuckBot and Slurp. No AI crawler is named anywhere, so all of it falls through to a near-empty User-agent: *. GPTBot and friends see none of that hygiene.

Finding 4 · the one to test first

The edge serves bots an empty page

From a datacenter IP, binance.com answers the homepage and /llms.txt with HTTP 202 and an empty body, and /robots.txt with 403. From a residential connection the same paths return real content. If that filter catches unallowlisted AI crawlers, no content strategy can work until it is fixed.

The thesis

The prize named in the role is definitional authority in AI-generated crypto content. Binance already owns the asset that wins it. Binance Academy is one of the largest crypto education libraries in existence and has its own sitemap index. It is absent from AI answers about the concepts it explains. Closing that is a retrieval and structure problem, and most of it is fixable without writing new content.

01

The citation probe

Seven buyer and learner questions run against the live retrieval layer that Perplexity, ChatGPT Search and AI Overviews draw citations from. Scored WIN 1.0 (Binance-owned page cited first), MID 0.4 to 0.5 (brand named in the answer, no owned citation), MISS 0 (absent).

QueryIntentResultScoreWho got cited instead
is Binance safe / proof of reservesbrandedWIN1.0binance.com/proof-of-reserves and Binance Square cited directly. The one clean win.
Binance vs Coinbase which is betterbranded comparisonMID0.5Every citation is third-party affiliate content: CoinLedger, CoinTracker, BitDegree, Datawallet, Benzinga. No Binance-owned page.
lowest fee crypto exchangecommercialMID0.4Binance fees quoted accurately, sourced from Kraken Learn, Koinly, Coingape. A rival's education hub is the authority on Binance's own pricing.
best crypto exchange for beginners 2026commercialMISS0.0Phemex, Coinbase, Kraken, Paybis, NerdWallet, Blockchain-Council.
what is staking crypto and how does it workdefinitionalMISS0.0Coinbase Learn, Kraken Learn and Crypto.com Learn all cited. Plus Revolut, N26, Britannica, Chainalysis. Binance Academy absent.
perpetual futures funding rate explaineddefinitionalMISS0.0Coinbase Learn, OSL, MetaMask, Messari, OneTrading. Binance runs the largest perpetuals venue in the world and is not the source.
how to buy bitcoin in Malaysiaregional transactionalMISS0.0StashAway, Transak, Changelly, LocalCoinSwap, BitDegree, fintechnews.my.
Prominence-weighted visibility by intent
Branded75
Commercial20
Definitional0
Regional0

Overall weighted index 27 / 100, n=7. Small sample, consistent direction.

Education-hub citations across the 5 unbranded queries
Coinbase Learn3
Kraken Learn2
Crypto.com Learn1
Binance Academy0

The pattern that matters. Rivals converted education hubs into the citation layer of AI answers.

Reading: the only thing separating a WIN from a MISS is whether the query names Binance. That is the signature of strong brand entity recognition sitting on top of weak content retrieval. Brand strength is already there. The retrievable answers are not.

02

The technical surface, six exchanges

Live-fetched. Reproduce any row with the tool I built for this audit ↗, which fetches these files at the moment you press the button.

Exchangellms.txtAI crawlers namedEntity graphRead
OKXpresent 15 sections, 73 links0Org + 11 sameAsThe most complete GEO surface of the six. Curated link directory, strong entity graph.
Coinbasepresent 6 sections, 24 links0Org, no sameAsTheir llms.txt actively routes agents to docs rather than marketing pages. Deliberate and well-judged.
KrakenHTML 4040Org + 12 sameAsStrongest entity graph and 29 sitemaps, no llms.txt. Still out-cites Binance on definitional queries through Kraken Learn.
Bybitpresent 13 sections, 39 links0Org, no sameAsExplicitly written for agents. One sitemap declared.
Crypto.compresent 5 sections, 0 links0no Org nodeProse guidance with no link directory, so agents get instructions and no destinations. Half-built.
Binancenone0not readableNothing served at that path. Entity graph could not be read from a datacenter IP (see 04).
The honest caveat

llms.txt is an emerging convention with no confirmed adoption by any major engine. Nobody should promise ranking from it. It costs almost nothing, it is how agents are increasingly told where to look, and being the only major exchange without one is a cheap gap to close.

Where the real leverage is

Not the file. Citation distribution studies consistently put Reddit, Wikipedia and YouTube far ahead of brand domains as cited sources. Entity presence and third-party surfaces move the number more than any owned-page change.

Nobody has solved this

Zero of six exchanges name a single AI crawler in robots.txt. The whole category is running AI crawler policy by accident. That is an opening.

03

The robots.txt non-inheritance trap

This is the finding I would open with internally, because it is mechanical, cheap to fix, and currently invisible.

What binance.com/robots.txt does today
User-agent: Baiduspider
User-agent: Bingbot
User-agent: DuckDuckBot
User-agent: Slurp
User-agent: Googlebot
Allow: /
Disallow: */my/
Disallow: */kyc-ui/
Disallow: */chart-webview/
Disallow: */backtest/
Disallow: */api/
...  52 disallow rules in this group

User-agent: *
Allow: */activity/challenge-competition/
Allow: */activity/collect-and-win/
Allow: */activity/referral-entry/
Disallow: */feed/sitemap/
What GPTBot actually reads

A crawler obeys exactly one group: the most specific one matching its token. Groups do not inherit and do not merge. GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended and CCBot are named nowhere, so every one of them falls through to User-agent: *.

The consequence: all 68 disallow rules protecting account pages, KYC flows, chart webviews, backtests and API paths apply to the six named classic crawlers and to nobody else. AI crawlers are, by omission, invited to spend their entire budget on the exact paths Binance already decided were not worth crawling.

The three Allow lines in the * group point at referral and competition activity pages. That is close to the only positive guidance any AI crawler currently receives from binance.com.

Fix

Add explicit groups for the AI crawler tokens, mirroring the hygiene already written for Googlebot, and add positive Allow paths for Academy, Support and Research. One file change, no content work, reviewable by Legal in a single pass.

Verify: the tool's second table counts, per domain, the disallow rules written for named classic crawlers that no AI crawler will ever read. OKX carries 39 such rules, Crypto.com 6.

04

The access question, and how I would settle it

Two vantage points, same paths, different answers. This is a hypothesis with a clear test, and I am presenting it as one.

Path on www.binance.comResidential IP (Malaysia)Datacenter IP (Cloudflare edge)What it suggests
/ (homepage)202 HTML interstitial202 empty bodyBot challenge, harder from datacenter ranges
/robots.txt200 3,888 bytes, full file403 520 bytesDatacenter egress is refused the crawl policy itself
/llms.txt202 empty202 emptyAbsent from both, so the missing-file finding holds
Why it matters

AI crawlers fetch from cloud ranges. If the WAF treats unallowlisted datacenter traffic as hostile, then GPTBot and PerplexityBot receive an empty 202 and index nothing. Any downstream GEO investment on affected paths would be spent on content those crawlers cannot retrieve.

Why I am not claiming it yet

My probes used ordinary browser and curl user agents from unallowlisted IPs. The published AI crawler ranges may well be allowlisted, and binance.com/proof-of-reserves does get cited, so some retrieval clearly works. A generic datacenter block is consistent with the evidence and not proven by it, and it cannot be the whole explanation for the citation misses.

The test, day one

Pull CDN and WAF logs, filter to the published IP ranges for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended, and produce the response-code distribution by path group. The answer is a single query away for anyone inside. If those crawlers are being challenged or 403'd, that becomes the highest-priority item in the role and everything else waits behind it. If they are being served normally, the problem is content structure and I move to section 05.

The Academy gap, the single biggest opportunity

The role names definitional authority as the goal. Binance already owns the corpus that should deliver it, and the corpus is not being retrieved.

The asset

Binance Academy has its own sitemap index in robots.txt, alongside Square, Research and Support. Thousands of explainer pages across every crypto concept, in dozens of languages. No competitor has a library this size.

The result

Zero citations across every definitional query tested. Coinbase Learn took three, Kraken Learn two. On funding rates, the venue with the deepest perpetuals book in the industry lost to Coinbase Learn and MetaMask.

Why rivals win it

Their pages answer in the first 40 words, sit on stable canonical URLs, carry visible authorship and dates, and are structured in chunks a retriever can lift whole. That is a formatting property, and it is learnable.

LeverWhat it changesWhy it moves citations
Answer-first rewriteOpen every Academy page with a 40 to 60 word standalone definition before context or promotion.Retrievers lift chunks, not pages. A chunk that answers alone is quotable. A chunk that needs the paragraph above it is not.
Chunk boundariesOne question per H2, self-contained under each heading, no cross-references that break when lifted.Chunk-level retrieval means the unit competing for a citation is the section, not the URL.
Entity scaffoldingOrganization and Article schema with author, dateModified and sameAs to Wikipedia, Wikidata and LinkedIn on every Academy page.Resolves Binance Academy as an authoring entity rather than an anonymous marketing page. Feeds E-E-A-T signals on AI surfaces.
Freshness signalsVisible dateModified and a real review cadence on the top few hundred concept pages.Engines weight recency on volatile topics. Undated crypto explainers get skipped in favour of dated ones.
Comparison pagesOwn Binance vs X and lowest fee exchange with honest, sourced tables.Today those answers are assembled entirely from affiliate sites with a commercial interest in the conclusion. A rival's Learn hub is currently the cited authority on Binance's own fee schedule.
Regional definitional pagesPer-market how-to-buy and regulatory-status pages, per licensed entity, in local language.Regional queries are the clearest misses in the probe, and the JD already calls for regional adaptation. Compliance-gated, so start where the licence is clean.
05

The eight duties, and how I would run each

DutyHow I would run itDetailed in
Own presence in AI-generated responses at scaleStart from a fixed prompt panel of 300 to 500 real questions, segmented by intent and market, scored on the WIN/MID/MISS index in section 01. Baseline first, publish the honest number, then move it.§01 §06
llms.txt, structured data, knowledge graph signals, regionalShip a root llms.txt pointing at Academy, Support, Research and the fee schedule, then per-region variants tied to each licensed entity. Organization plus sameAs on the root, Article plus author plus dateModified on Academy.§02 ★
Define GEO as a channel with impressions, citations, attributionThree tiers so the number survives scrutiny: citation share from the prompt panel, crawler behaviour from CDN logs, referral sessions from AI hosts. Never blend them into one vanity figure.§06
Translate citation gaps into content requirementsEvery MISS becomes a brief naming the query, the current winner, the chunk that beat us and the specific structural change. Content receives requirements, not opinions.
Ongoing prompt testing and regression catchingWeekly automated panel re-run into a time series. Alert on citation loss for defended queries. Treat losing a citation like losing a ranking.§06
Benchmark Coinbase, Kraken, OKX, Bybit, Crypto.comAlready started, section 02. Same probe run for all six, same scoring, tracked as share of voice rather than a one-off snapshot.§01 §02
Log-level and query-level analysis of AI crawler behaviourFirst job is the access question in section 04: response-code distribution for published AI crawler ranges by path group. Then crawl frequency against citation outcomes to find which page groups are fetched and never cited.§04
Track trends and turn them into frameworks for the SEO teamI already maintain a working reference on retrieval mechanics, citation triggers, query fan-out and entity salience. Internally that becomes a short living playbook plus a monthly note, written for people who ship rather than for a conference talk.§06
06

Measurement, built where none exists

The role says the framework does not exist yet. Whoever defines it sets what the company believes about this channel, so it has to be honest before it is flattering.

TierMetricSourceWhat it is good and bad for
1 · CitationWeighted citation index and share of voice per intent segment and marketFixed prompt panel re-run weekly across ChatGPT, Perplexity, Gemini and ClaudeThe only tier that measures the actual objective. Costs API budget and moves with model updates outside our control.
2 · RetrievalAI crawler hit rate, response codes, crawl depth by page groupCDN and WAF logs filtered to published crawler IP rangesLeading indicator and the only tier that is fully ours. Says nothing about whether we get cited.
3 · OutcomeSessions, signups and funded accounts referred from AI hostsReferrer analytics on chatgpt.com, perplexity.ai and similar, plus a dedicated channel groupingThe number leadership will care about. Small today, undercounts badly because many assistants strip referrers.
Two rules I would set on day one

Never report tier 3 alone, because referrer stripping makes it look like nothing is happening. Never report tier 1 alone, because it is measurable without being bankable. The pairing is what makes the channel legible to a CFO.

Build or buy

Buy a commercial tracker for coverage and continuity. Self-host the panel in parallel so the time series is owned, reproducible and available before a vendor dashboard refreshes. I have run both.

07

First 90 days

Sequenced so the cheap reversible changes ship while the slow ones clear review. Nothing here assumes authority I would not have.

Days 1 to 30 · establish truth
  • Settle the access question from CDN logs. Everything else is downstream of the answer.
  • Stand up the prompt panel, 300 to 500 queries, and publish the baseline even if it is bad.
  • Map the fifteen sitemap groups to page templates and find which carry schema.
  • Meet Content, Tech, Data and Compliance. Learn the actual review path and real ship times before promising anything.
  • Draft the robots.txt AI crawler groups. One file, one review pass.
Days 31 to 60 · ship the cheap wins
  • Ship robots.txt AI groups and root llms.txt. Low risk, fully reversible.
  • Organization and sameAs on the root, Article with author and dateModified on the Academy template.
  • Answer-first rewrite of the 50 Academy pages behind the highest-value definitional misses.
  • Competitor share-of-voice tracking live for all five named rivals.
  • First monthly note to the SEO team on what moved and what did not.
Days 61 to 90 · make it a channel
  • All three measurement tiers reporting on one dashboard with an agreed definition.
  • GEO requirements embedded in the content brief template so new pages ship correct by default.
  • Regional llms.txt for the two or three cleanest licensed markets.
  • Entity work off-site: Wikidata and Wikipedia accuracy, and the third-party surfaces that dominate citation distribution.
  • Second baseline re-run. Report the delta honestly, including anything that got worse.
What I would not promise

A citation-share number by day 90. Model updates move these results independently of anything we ship, the sample needed to detect a real change is larger than one quarter provides, and Compliance review time on a regulated exchange is not mine to control. What I would commit to is a defensible baseline, the technical gaps closed, and a measurement system the team keeps using after the novelty wears off.

08

Method & sources

How the probe was run

Seven questions run against the live web retrieval layer that Perplexity, ChatGPT Search and AI Overviews draw citations from, on 3 Aug 2026. Scored WIN 1.0, MID 0.4 to 0.5, MISS 0, then weighted per intent segment. Same method as an earlier audit I ran at n=19 on a different brand.

Limits, stated plainly

n=7 shows a direction and is too small to be a measurement. This is the shared retrieval layer rather than four separate engine APIs, so per-engine share of voice is not broken out. All requests came from Malaysia or the Cloudflare edge, and binance.com geo-restricts Malaysia, so the HTTP status codes in section 04 are vantage-dependent by design. The llms.txt absence was confirmed from both vantage points.

Sources

Independent analysis of public data, prepared as an application work sample. Not affiliated with, endorsed by, or produced for Binance or any exchange named here. Findings reflect what those properties served on 3 Aug 2026 and will change as they do.

Try the live GEO readiness tool ↗