Two days ago I published a census of every AI crawler that hit my site over 34 days. The headline was that ChatGPT now fetches my pages for live users more often than Googlebot crawls them. This post is about the number in that dataset I couldn't stop thinking about — the one that isn't ChatGPT at all.
It's bingbot.
The ratio nobody is optimizing for
Fresh numbers from this morning's run (36 days of Caddy access logs, HTTP 200 responses only, for a small business site):
| Bot | Hits | Per day |
|---|---|---|
| bingbot | 4,488 | 123.3 |
| ChatGPT-User (live retrieval) | 1,394 | 38.3 |
| Googlebot | 1,073 | 29.5 |
| GPTBot (training) | 505 | 13.9 |
| OAI-SearchBot | 249 | 6.8 |
Bing's crawler reads my site 4.2 times more than Google's. For years I would have read that as wasted crawl budget from the search engine nobody uses — Bing's share of human search traffic barely registers in my analytics.
Then you put the second fact next to it: Seer Interactive ran 500+ citations through SearchGPT and found that 87% of them matched Bing's top-10 organic results for the same query. Only 56% matched Google's top 10.
ChatGPT's web search doesn't have its own index. When it goes looking for sources to cite, it queries Bing. Those 4,488 bingbot hits aren't for Bing's ten human users — they're the supply chain for every ChatGPT answer that might mention you. Bing stopped being a search engine and became an API for AI answers, and its crawler's appetite finally makes sense.
What the side door looks like from inside the logs
The chain is visible end to end in my own data. bingbot's most-crawled pages this window are my blog's long-form posts and guides. And the pages ChatGPT-User actually fetches when a human asks something? The same category: my Chatwoot-vs-Intercom comparison got 101 live ChatGPT retrievals in 36 days, and a Hebrew post about WhatsApp spam law got 226 retrievals from Claude-User alone. Comparison pages and definitive guides go in through Bing, come out as citations.
I run a WhatsApp automation studio in Israel, and this pipeline stopped being theoretical for me in mid-July, when a Google core update cut my Israeli search clicks nearly in half. The traffic that kept arriving through the collapse was exactly this AI-assistant channel — which, per Seer's data, is largely downstream of an index whose webmaster console I had never once opened.
One more uncomfortable detail from the earlier post: when I compared server logs to Google Analytics, GA4 saw about 11% of the visitors the logs saw. AI-referred and privacy-shielded traffic is mostly invisible to your dashboard. If you only look at GA4, this entire channel doesn't exist.
The practical checklist
If 87% of ChatGPT citations come from Bing's top 10, then "rank in Bing" is now a growth channel with almost no competition, because everyone else is still fighting over Google. What I've changed:
1. Stop treating bingbot as a nuisance. Plenty of robots.txt files and WAF rules throttle or block bingbot to "save crawl budget." That's now equivalent to blocking your ChatGPT visibility at the front door. Check yours before anything else.
2. Register in Bing Webmaster Tools. It takes minutes (you can import your verified site straight from Google Search Console) and it's the only place you'll see how Bing actually indexes you.
3. Ship IndexNow. Bing supports instant push indexing — a tiny API ping on every publish instead of waiting for a crawl. For a channel where citations follow the index, freshness is free ranking.
4. Feed it the content type the pipeline rewards. In my logs, what gets retrieved for live users is comparisons ("X vs Y"), pricing breakdowns, and complete guides — pages that answer a full question. Listicle fragments don't get cited.
5. Measure in your logs, not your analytics. Thirty seconds on any server:
# bingbot vs Googlebot, raw hit counts
grep -c 'bingbot' access.log
grep -c 'Googlebot' access.log
# which pages bingbot actually wants
grep 'bingbot' access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head
(Adjust the $7 field for your log format; mine is Caddy JSON piped through a small parser.)
The uncomfortable conclusion
I spent years doing SEO for an audience of humans on Google. The team behind 50+ WhatsApp bot deployments that pays my bills gets found today by a different reader: a retrieval bot working for an AI that a buyer asked in natural language. That reader enters through Bing, and almost nobody is holding the door.
Update (19 Aug): I checked, and part of this post is wrong
A commenter, @bulti_global, pushed back on turning the 87% overlap into "citations follow the Bing index," and asked a sharp question: are there pages that rank in Bing's top 10 but never trigger a retrieval? He suggested splitting the funnel page by page instead of trusting the aggregate bot ratio.
Answering it forced me to open Bing Webmaster Tools — where I found an account I'd forgotten I had, holding 91 days of data. Two corrections follow from that.
First, I overstated my own ignorance. I wrote that this ran through "an index I had never once logged into." I had registered; I had just never looked.
Second, and worse for my thesis: at page level, Bing rank does not predict AI citation. Same site, same 91-day window (19 May – 17 Aug). Bing organic delivered 650 impressions and 4 clicks. AI citations over that window: 81.
| Page | Bing impressions | Avg position | AI citations |
|---|---|---|---|
| whatsapp-bot-platforms-comparison-2026 | 169 | 4.48 | 62 |
| whatsapp-automation-guide | 90 | 4.00 | 0 |
| whatsapp-bot-for-business | 60 | 4.20 | 0 |
| free-whatsapp-bot-guide | 46 | 5.87 | 7 |
| pricing | 31 | 4.32 | 0 |
| whatsapp-bot-clinics | 29 | 3.28 | 0 |
| whatsapp-bot-for-gyms | 16 | 2.81 | 0 |
My best-ranking page in Bing — average position 2.81 — has never been cited. Neither have 3.28, 4.00 or 4.32. The page that takes 62 of the 81 citations ranks worse than four pages that were never cited once, and one page picked up 7 citations without appearing in my top-14 by impressions at all.
What separates the winner looks like format, not position: it is the only head-to-head "X vs Y vs Z" comparison table on the site. Item 4 of the checklist below survives. The implicit "rank in Bing and citations follow" does not — not at page level, not on this sample.
One caveat on the numbers, because it matters: Bing labels that report "Microsoft Copilots and Partners." It measures the Microsoft AI surface, not ChatGPT specifically. It is adjacent to the claim, not identical to it.
Registering still turned out to be the single highest-value thing in this post — just not for the reason I gave. It is where you find out that your best-ranking page is invisible to the answer engines.
Run the grep above on your own logs and tell me your bingbot-to-Googlebot ratio in the comments — I'm collecting data points for a follow-up post. Mine is 4.2:1. If yours is under 1:1, I'd especially like to know what kind of site it is.
Top comments (15)
This is genuinely interesting, and I love that you left the correction in the article because i was already going to brag on this insight to my team. The updated result is more interesting than the original claim and thankfully more correct. Bing rank starts to look less like the destination and more like an entry ticket; very interesting to see that the presentation and content matter more then mere ranking , maybe it's looking for non obvious clues
I wonder whether a useful next experiment would be a matched-pair test: take two pages targeting similar intent, give one a clear comparison table and concise evidence blocks, leave the other unchanged, then watch retrievals and citations over several weeks. That could help separate “the format is causal” from “comparison queries simply get cited more.” Very new territory, but that distinction feels worth chasing.
Thanks — and the matched-pair design aims at exactly the confound I'm stuck on.
The trouble is my own sample can't settle it. The page carrying 62 of the 81 citations is a comparison, and comparison questions are also the ones people bring to an AI assistant in the first place. So "the format is causal" and "comparison intent simply gets cited more" are perfectly collinear in my data — no amount of re-reading the existing logs pulls them apart.
Two things I'd add to your design if I run it. First, publish both pages together and change one of them later, using the unchanged page as its own before/after; otherwise publication age quietly becomes the variable you're measuring. Second, score it on citations rather than impressions, because the whole finding was that rank and citation came apart at the page level: the position-2.81 page has zero citations, while a page that never cracked the top 14 by impressions earned seven.
If I get the pair up I'll post the numbers either way, including the boring outcome where the format does nothing.
This is exactly the kind of boring experimental discipline that makes the result interesting. The comparison-intent confound is a knot you can't untie by staring harder at the same data. Publishing the pair together and then changing one page later is smart—it turns the unchanged page into a little time machine.
I'd be especially curious whether the citation effect appears first in fresh answers before it shows up in search impressions, or whether both move together. If you run it, even a null result would map the terrain.
Don't know yet, and that ordering is the measurement I most want. The awkward part is that the two signals aren't sampled the same way, so "which moved first" is partly an artifact of how often I can look.
Search impressions arrive daily and pre-aggregated — I get them whether or not I was paying attention. Citations only exist at the moment someone asks. There's no impression log for them, so the only way to observe one is to ask on a schedule and record what comes back, which makes my probe interval the floor on how fast I can detect any change at all. If I probe weekly and impressions update daily, impressions will look like they moved first every single time, regardless of what actually happened.
So the answerable version is narrower than the question: does the citation change show up in the first scheduled probe after the edit, while impressions are still flat for several days after? That requires a probe cadence tight enough to distinguish "moved first" from "I looked sooner" — daily probes against a fixed query set, timestamped against the edit, and the edit made on a day nothing else on the site changes.
Agreed on the null result, and the pair is the reason it's worth running rather than a nice-to-have. If both pages move together, the effect wasn't the edit — it was something seasonal or index-wide. That's exactly the outcome I'd otherwise write up as a success, which is the failure mode the unchanged page exists to catch.
Love how carefully you're separating “moved first” from “I happened to look first”—that distinction is where a neat experiment becomes a real one. The sampling clocks are basically two cameras recording at different frame rates, so chronology needs its own calibration.
Would a staggered cadence help: daily citation probes around the edit window, then taper to weekly once impressions respond—or use a few fixed query families so one noisy prompt doesn't impersonate a trend? Your unchanged twin page still feels like the hero here; it keeps a good story from outrunning the evidence.
Staggered cadence is the right shape, with one catch: the taper has to be tied to the edit clock, not to when impressions respond. If I drop from daily to weekly the moment impressions move, the sampling rate becomes a function of the outcome — a citation change that arrives late gets detected late for a reason that has nothing to do with the page. So: daily for a fixed window after the edit, decided in advance (I'd take three weeks, since Bing's refresh on my pages has run one to three), then weekly.
Query families — yes, and they fix a second problem I hadn't named. A single prompt can't separate "the page got cited" from "this particular wording happened to hit it". Three or four families (direct comparison, a price question, "best X for Y", plain informational), each with a few phrasings, scored as the fraction of phrasings that cite the page. One noisy prompt then moves a fraction, not a binary. And the twin gets the identical families on the identical schedule, or it's a control in name only.
The part I still don't have a design for is the answer surface itself. The engine changes under both pages at once — retrieval, model, whatever Microsoft ships that week — and the twin cancels that only if both pages are equally exposed to it. Two comparison pages probably are; a comparison paired with a guide isn't. That's the clause of "similar intent" I'd be strictest about when picking the pair.
That fixed-window point is exactly the guardrail I was missing—the sampling plan needs to be independent of the outcome, or the measurement starts chasing its own shadow. Three weeks daily and then weekly feels defensible if the window is chosen from observed crawl-refresh lag rather than impressions.
The answer-surface confound may be the harder beast. Could you add a small panel of matched twins across page types and intent clusters, then treat engine-wide changes as a shared time effect? It would not remove every confound, but it could show whether one lucky pair is carrying the conclusion.
A panel is the right move, and it does more than average out a lucky pair — it turns the "engine changed under both pages" problem from a confound into a term I can estimate. Every pair sees the same week, so the week is a shared effect; what's left after removing it is what the edit did. One condition for that to work: the pairs cannot all be edited at once. If every twin flips in the same week, the edit and the week are the same variable and the model has nothing to separate them with. So the edits get staggered across pairs on a schedule fixed in advance — which is the same rule as the taper, applied one level up.
The honest constraint is scale. This is one site, and 62 of 81 citations sit on a single page; most pages have zero. A panel here is four to six pairs, not forty. At that size I won't pretend to a significance test. The check I can run is leave-one-pair-out: re-fit with each pair dropped, and if the sign of the edit effect survives every drop, no single pair is carrying it. If it doesn't, that's the result, and I'll say so.
On the window: agreed that it comes from crawl-refresh lag, with one caveat. The lag I've measured is fetch-side (when Bing re-crawled the page). Whether a citation moves lags that further, and I don't yet have a number for it. So the pre-registered window is crawl-to-first-citation-move on an untreated page — measured once, before any edit, then fixed.
I'll publish the pair list and the staggered schedule before the first edit, so the design can't drift toward the outcome afterward.
I like that you're putting the pair list and schedule out before the results—much more interesting to follow an experiment unfolding than another tidy success story after the fact. And the small-site constraint makes this feel relevant to people who don't have thousands of pages to play with.
The practical question I'm curious about is whether the edit helps an uncited page get its first foothold, or mostly helps pages already being picked up. Those would lead me to quite different writing decisions. Are you hoping the pairs will give you a glimpse of that distinction too?
That's the split I'd most want to see too, and my own site settles most of it before the experiment starts. Almost every page here is in the uncited bucket; the page holding 62 of the 81 citations is the exception, not the top of a gradient. So a four-to-six pair panel is five foothold pairs and, if I'm lucky, one already-cited pair. Anything I claim about the second case would rest on that one pair, which is an anecdote wearing a table.
The two questions also cost different amounts to measure. A first foothold is close to binary: the first probe that cites the page has already answered it. For a page cited in, say, half the phrasings, the edit has to move a fraction, and telling 0.5 from 0.65 apart needs more probes per week than I can run.
So the honest scope is one arm, not two. The panel can speak to the foothold question. For the other I'll report what the single pair did and label it as one pair, rather than let it ride along as if the design covered it.
The 4.2:1 crawl ratio is a useful operational signal, but I’d be careful turning the 87% overlap into “citations follow the Bing index.” In a separate 100-query ChatGPT shopping test we ran, 75% of the top products matched Google Shopping’s top 3. Different surface and intent, but that contrast suggests the retrieval supply chain may vary by mode. For the follow-up, I’d split the funnel page by page: Bing index/rank → ChatGPT-User fetch → actual citation. Have you found pages that rank in Bing’s top 10 but never trigger ChatGPT-User? Those misses may be more diagnostic than the aggregate bot ratio.
Following up, because I was wrong about my own setup and it turned out to matter.
I said I had never registered in Bing Webmaster Tools. I had — the account was sitting there with 91 days of data I had never once opened. So I can actually answer your question now.
Caveat first: BWT labels that report "Microsoft Copilots and Partners," so it measures the Microsoft AI surface, not ChatGPT specifically. Adjacent to my claim, not identical to it. Same site, same 91-day window (19 May – 17 Aug): Bing organic gave 650 impressions and 4 clicks. AI citations over the same window: 81.
The page-level funnel you asked for:
You called it. My best-ranking page in Bing — position 2.81 — has zero citations. So do 3.28, 4.00 and 4.32. The page taking 62 of the 81 citations ranks worse (4.48) than four pages that were never cited once. And
business-automation-guidepulled 7 citations without appearing in my top-14 Bing pages by impressions at all.So on this site, at page level, rank does not predict citation. The misses were more diagnostic than the aggregate ratio, exactly as you said. What separates the winner looks like format rather than position: it is the only head-to-head "X vs Y vs Z" comparison table in the set.
One thing your funnel split surfaced that I had not considered: citations were flat zero until 23 June, then 25 in the first 30 days and 56 in the last 30. Bing impressions did not move like that at all. Whatever gates citation is not stepping in time with the index.
Fair push, and it makes me separate two things I ran together.
The 87% isn't my measurement — it's Seer Interactive's, over 500+ SearchGPT citations. My own data is just the crawl ratio and the per-page fetch counts. So your shopping result doesn't really contradict it, it bounds it: different surface, different intent, plausibly a different retrieval path. I generalized a SearchGPT-shaped finding into "AI citations" and should have marked that seam.
On your direct question — no, I can't answer it yet, for a reason that's awkward given item 2 of my own checklist: I have no Bing rank data. I never registered in Bing Webmaster Tools. I have the fetch column of your funnel and nothing for the index/rank column in the middle, which is exactly where the misses would show up.
What the fetch side does say is that it's brutally concentrated. One Chatwoot-vs-Intercom comparison took 101 of 1,394 ChatGPT-User hits in 36 days, and most of the site got zero. That shape argues for your point rather than mine: an aggregate bot ratio averages over a large tail of pages that never enter the funnel at all, so it can look healthy while the per-page reality is a handful of winners.
Registering in BWT is now the blocking step for the follow-up. Page-by-page, misses first.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.