How SERP Snippet Indexing Drives AI Search Visibility
TL;DR: We published five fresh pages, blocked Bingbot on every one of them, and ChatGPT still cited them within hours of Google indexing. The only public surface left was Google's SERP. So if you win the Google snippet, you raise the odds an answer engine repeats your content, even when Bing hasn't caught up. Tighten your answer block, ship schema, push freshness, and keep GPTBot unblocked. That's most of the AI-search playbook.
The punchline first: a Bing-blocked page got cited anyway
Ask most SEOs how ChatGPT finds a brand-new page and you get a two-word answer: Bing crawl. OpenAI's deal with Microsoft hands the model Bing's index, so the logic is clean. That was my assumption going in, until I ran a test designed to break it.
I published a fresh URL, added User-agent: Bingbot Disallow: / to robots.txt, and confirmed Bing returned zero results for the path. I pushed the page through Google Search Console's URL Inspection tool and waited. Inside three hours it showed up in Google's top-20. Then I asked ChatGPT's live-web mode a question the page was the obvious answer to, and it cited the page. Bing still had nothing. The only public index that knew this URL existed was Google's.
One run is an anecdote, so I did it four more times. Same shape every time: page blocked from Bing, indexed only by Google, cited in ChatGPT anyway. The citation usually appeared within a few hours of Google indexing. The fifth run took the longest, but the citation still landed before any Bing entry did.
I want to be honest about the size of this. Five URLs is a small sample, and I'd happily watch someone run it across fifty domains and tell me I'm wrong. But the direction is hard to wave away: when ChatGPT's own crawl and Bing both come up empty, it appears to reach for Google's snippet. That's the finding the rest of this piece hangs on.
It also pushes back on the advice that's circulated since browsing rolled out: "optimize for Bing because of OpenAI." If Google's SERP can seed a ChatGPT answer, optimizing only for Microsoft's index leaves coverage behind.
How the test was actually run
For anyone who wants to replicate it, here's the exact procedure. Nothing clever, which is the point.
- Published five brand-new URLs over two weeks, one at a time, each on a domain ChatGPT had never cited before.
- Added
User-agent: Bingbot Disallow: /torobots.txton every URL, then verified Bing returned zero results for the path. - Submitted each URL through Google Search Console's URL Inspection tool and waited for Google to index it (median: under three hours).
- Queried ChatGPT's live-web mode with a question the URL was the obvious answer to, and recorded whether the page was cited.
- Re-ran the same query 24 hours later to confirm the citation persisted, rather than being a one-off cache hit.
The Bingbot block is the whole experiment. It removes Bing as an explanation, so if the page still gets cited, Google's SERP is the only public surface left standing. Aleyda Solis ran a related test in July 2025, showing ChatGPT answering from a Google snippet on a page Bing didn't have, and Wellows ran an indexed-versus-unindexed version. Our addition is the explicit Bing block. I'll call ours a parallel replication, not the first; Solis published before I wrote this up.
A three-layer model for how live-web mode appears to work
Here's where I slow down, because this is the part people quote without the qualifier. What follows is a model I built from watching behavior, not from anything OpenAI has documented; treat it as a working hypothesis, not a spec sheet. When you ask ChatGPT for fresh information, it doesn't fire a single crawler. It seems to follow a fallback chain with three layers, and the chain shows surprising deference to Google.
Layer 1: OpenAI's on-demand fetch. The browser tool spins up a live fetch via GPTBot for the handful of URLs it predicts are relevant. In my own timing, the page became citation-eligible somewhere in the range of two to seven minutes after the fetch fired. I want to be clear that this number is mine, inferred from response timing, not a figure OpenAI publishes. The phrase "ephemeral index" is my shorthand for what seems to be happening, not OpenAI's terminology. Read it as a hypothesis.
Layer 2: Bing index lookup. If the live fetch times out or returns thin content, the model appears to query Microsoft's Bing API. OpenAI runs on Azure, which Microsoft also uses for Bing, so I'd guess this lookup is low-latency, though OpenAI hasn't published any figures and I can't measure it directly. It's also bounded by whatever Bing already knows, which is often a narrower slice of the web than Google.
Layer 3: external snippet fallback. This is the interesting one, and the one the Bing-blocked test points at. When neither the live crawl nor Bing has the page, ChatGPT seems to reach into public indexes it doesn't own. Google's SERP, mostly. It reads the rendered result page, lifts the snippet text, and treats that as a cached summary it can cite. That would explain how a Google-only, Bing-blocked page surfaces in an answer within hours.
I'm spelling out the uncertainty on purpose. I inferred Layer 3 from behavior, not documentation. There could be a fourth mechanism I'm not seeing, or OpenAI could change the chain next month. But the Bing-blocked result is hard to explain any other way, and that's enough to act on.
Page signals that seem to govern whether you get cited
Early experiments (mine and others') point at a handful of page attributes that tilt the odds. Some are well-established Google signals; a couple are my own observations, flagged as such. The right column is the move that helps both your Google snippet and your AI-answer odds.
| Signal | Why it seems to matter | What to do |
|---|---|---|
| Snippet-ready answer block | Google uses it for the SERP description, and that same block is what gets lifted when the Layer 3 fallback triggers. | Put a 40 to 60 word TL;DR under the H1 with the primary query phrase. |
FAQPage and HowTo schema |
Structured data feeds Google Featured Snippets, which are clean, self-contained scrape targets. | Add JSON-LD FAQs and validate in the Rich Results Test. |
| Fresh index timestamp | Hypothesis, not fact: in my testing, recently-indexed pages seemed to win "latest" prompts more often (loosely, inside ~48 hours). I can't prove a hard recency cutoff. | Ping the Indexing API or GSC URL Inspection right after publish. |
| Open AI-crawler access | If GPTBot hits a 4xx, Layer 1 fails and you're relying on the lower layers to save you. | Keep robots.txt at User-agent: GPTBot Allow: /. |
| Semantic heading hierarchy | Descriptive H-tags help both Google and the way models chunk a page into answerable pieces. | Write H2/H3s that name the sub-topic; skip the generic "Conclusion." |
| Low boilerplate | Repetitive intros get down-ranked, which lowers the chance the page is chosen for a snippet at all. | Trim the fluff. Get to the unique value inside the first 100 words. |
| Engagement (indirect) | I have no evidence ChatGPT reads time-on-page directly. What I'd argue is the chain rank → citation: better engagement tends to lift rank, and rank correlates with citation odds. | Improve LCP/INP, embed visuals, add internal links so readers keep exploring. |
The freshness row and the engagement row are inferences, and I'd rather label them than smuggle them in next to Google's published thresholds. If you only trust the unflagged rows, you'll still do most of the useful work.
What the AI Overviews data does and doesn't tell us
This is the section I most want to get right, because it's where I see the most confused thinking, including in my own earlier drafts. There's a popular dataset I almost misused. Passionfruit analyzed over a million AI-generated summaries and found that 40.58% of citations come from Google's top 10, with citation probability stratifying sharply by rank: 33.07% at position #1, dropping to 13.04% at position #10. Striking gradient. Easy to grab as proof that "AI loves your Google rank."
The catch: those numbers describe Google AI Overviews, not ChatGPT. And that distinction matters, because AI Overviews is Google's own product pulling from Google's own index, so a tight correlation with Google rank is almost circular. It tells you Google trusts Google. The same Passionfruit analysis is blunt about ChatGPT, finding only about 6.82% overlap between ChatGPT's results and Google's top 10. That's nearly the inverse of the AI Overviews story.
So I'm not going to claim Passionfruit proves my Layer 3 hypothesis. It doesn't. It shows AI Overviews leans hard on Google rank (unsurprising) and that ChatGPT diverges from Google's top 10 most of the time. My five-URL test is a separate, much smaller piece of evidence about one corner case: what ChatGPT does when its own crawl and Bing both miss. Two different questions. Stapling the Passionfruit numbers onto the ChatGPT claim was exactly the move I had to stop myself from making.
Where that leaves us is more modest but more honest. For AI Overviews, your Google rank is close to the whole game; the data is strong and the mechanism obvious. For ChatGPT, Google rank is one input among several, and the snippet fallback I observed looks like a backstop that fires when the faster layers come up empty, not the primary path. Two engines, two stories. Don't let one set of numbers do double duty.
A few common misreadings flow from collapsing that distinction:
- Treating AI search as Bing-only. If you optimize only for Bing because of the OpenAI deal, you skip the Google snippet path entirely. My Bing-blocked URLs still got cited; the page-level lever was Google snippet eligibility, not Bing speed.
- Treating AI Overviews as a separate ranking system. Gemini-powered AI Overviews pull from the same top-of-page snippets humans see. There's no hidden "AI rank" underneath. Pages outside the top 10 are rarely cited there.
- Treating every AI citation as one black box. AI Overviews citations track Google rank tightly. ChatGPT citations don't, by the same dataset. If you assume a single rule, you'll over-invest in the wrong lever for one of the two.
The unflashy takeaway: a lot of AI-visibility work is just classic snippet work done one notch sharper, and it holds whether the engine is Google's overview or ChatGPT reaching for a fallback.
Optimizing for two indices without doubling the work
It's worth treating Google and Bing as complementary feeds into a single answer surface rather than two channels. Classic blue-link traffic still arrives through both. But ChatGPT's live-web mode, at least in my testing, seems to lean on Google's snippet when its own crawler or Bing lags. Abandoning Bing would still be a mistake: Bing's API is ChatGPT's first stop in the chain, and Copilot's SERP keeps growing its own traffic. The play isn't "pick one." It's to make the page legible to both, which mostly happens for free if you do the snippet work well.
| Task | Wins Google snippet (Layer 3) | Wins Bing (Layer 2) | Action |
|---|---|---|---|
| Indexing API / URL Inspection | Forces near-instant snippet eligibility (I saw under three hours). | No equivalent for non-jobs content. | Push the API call for high-value posts; watch the "Crawled" timestamp. |
FAQPage / HowTo schema |
Strong correlation with Featured Snippets. | Bing shows FAQ drop-downs, lifting CTR. | Add concise Q&A pairs; validate in Rich Results and Bing Webmaster Tools. |
| Answer block under 60 words | Google uses it for the meta snippet; it's the text Layer 3 lifts. | Fits Bing's ~160-char snippet cap. | Place it below the H1 with the target query phrase once. |
| Allow reputable AI crawlers | If the Google snippet path fails, ChatGPT falls back to a GPTBot fetch. | Same for Copilot and Perplexity. | robots.txt: User-agent: GPTBot Allow: /, plus Google-Extended. |
I've started treating snippet fitness as a KPI alongside rank and clicks. If a page can't win a Google Featured Snippet, I don't expect it to win prime space in an AI answer either.
What could break this, and what won't
I don't want to oversell a model built on five URLs and a fallback I inferred, so here's what I'm watching. Google could throttle large-scale snippet scraping at any point, through rate limits, obfuscated HTML, or a paid API, which would force OpenAI back onto its own GPTBot. On the other side, OpenAI keeps scaling its crawler fleet, and a genuinely comprehensive proprietary index would dilute Google's role in the chain. Either move would change Layer 3, maybe erase it.
There's also Google's Web Guide experiment, which I dug into in our Web Guide optimization guide. By grouping URLs under AI-generated headings, it may change which snippets, and how many, are exposed for any external agent to read. A page that slips from a clean top-10 slot into an expandable bucket could get harder to scrape. And the fair-use fight is simmering under all of it: publishers want compensation or opt-out, regulators are watching, and a legal precedent could redraw what "public" means for SERP data.
I don't have a confident prediction about which lands first. What I'm confident about is narrower: clean markup, fast pages, and snippet-ready content stay valuable no matter how the plumbing shifts. That's the part I'd build on even if the rest moves under me.
Your Google game still feeds AI answers
Bing optimization remains useful for plain blue-link visibility, but it isn't full coverage for AI search. The evidence I have (a five-URL experiment plus behavior I keep seeing across the sites we work with) points to Google snippets acting as an unofficial feed for ChatGPT when the faster layers miss. Control what Google surfaces and you raise the odds the model echoes you, even while Bing's crawler is still catching up.
So tuning answer blocks, schema, and freshness for Google isn't only traditional SEO anymore. In 2026 it's also how you show up in AI answers, and the people who internalize that this is one dual-index workflow get to be early in citations before the rest of the field notices.
Run the free AI Visibility Checker to see whether your snippets are getting picked up by ChatGPT, Gemini, and Perplexity, and where you're getting skipped.
Frequently asked questions
Does blocking GPTBot hurt AI search visibility?
In my tests, yes, but indirectly. Blocking GPTBot only kills Layer 1, the on-demand fetch. Layer 2 (Bing) and Layer 3 (the Google snippet fallback) still operate, so a page with strong Google rankings can still surface in ChatGPT citations even with GPTBot disallowed. Allowing GPTBot is the cheapest way to cover all three layers and avoid the edge case where Bing and Google are both stale.
How fast does Google indexing show up in ChatGPT citations?
In my five-URL replication, the page appeared in ChatGPT live-web answers within a few hours of Google indexing, with one slower outlier. I push the Indexing API or URL Inspection on publish to compress the first half of that window. The rest depends on ChatGPT's snippet-cache refresh, which I can't influence directly. Treat the timing as an observation from a small sample, not a guarantee.
Should I still optimize for Bing if Google snippets feed ChatGPT?
Yes. Bing is still ChatGPT's first stop in the fallback chain, so a Bing-indexed page reaches Layer 2 before anything else fires. Bing also powers Copilot's SERP, which has its own growing traffic. The point isn't to abandon Bing; it's to stop treating Bing as the only AI-search lever. Optimize for both indices and you cover Layer 1 through Layer 3.
Do the Passionfruit citation stats apply to ChatGPT?
No, and this trips people up. Passionfruit's 40.58% / 33.07% / 13.04% rank figures describe Google AI Overviews, which pull from Google's own index, so the correlation with Google rank is close to circular. The same analysis found only about 6.82% overlap between ChatGPT's results and Google's top 10. Use those numbers to reason about AI Overviews, not ChatGPT.
Is the Bing-blocked finding reproducible?
Aleyda Solis ran a related test in 2025 showing ChatGPT pulling from Google SERP snippets, and Wellows replicated a narrower indexed-versus-unindexed version. My addition is the explicit Bingbot block, which isolates Google as the only public source. Five URLs is a small sample, and I'd genuinely welcome larger replications across more domains and prompt types.
Keep reading
- Optimising for Featured Snippets: the snippet-extraction signals AI engines reuse.
- Google Web Guide Optimisation: how Gemini reorganises results into AI-generated buckets.
- AI Visibility Checker: see whether your snippets are being picked up by LLMs.
Read More
no credit card required