ChatGPT Snippets Are the First ~200 Characters of Your Body Text

Conclusion

We ran five prompts through ChatGPT and pulled 383 search-result snippets from the conversation JSON. Every one was 200–206 characters long, 100% were truncated at the end, and on all six pages we checked, not a single character of the meta description appeared.

So what is being cut? We fetched the actual pages and compared them. On all 12 pages we could match, boilerplate such as the header, global navigation and "Skip to content" had been discarded. It isn't a mechanical 200 characters from the top of the page; it's the first 200 characters of the extracted body text.

When the same URL is picked up by a different query, the snippet doesn't change by a single character. Across days, however, 18% changed, and in some cases body extraction failed and navigation filled the slot.

In short, what ChatGPT receives is the first 200 characters of your body text. What lands there shapes how your company gets introduced.

All lengths in this article are character counts, not word counts, taken from a Japanese-language study whose sample includes both Japanese and English pages.

Figure

What 383 snippets showed

383

snippets

Unique count after deduplication

100%

were 200–206 characters

Median 202 characters; all truncated at the end

0

contained meta description text

Across all 6 pages checked

82%

unchanged across days

73 of the 89 that appeared on both days

What we looked at

ChatGPT's conversation JSON records the search queries the model issued and the URLs, titles and snippets that came back. This article deals with the search_result_groups[].entries[].snippet field.

What we looked at	Where it is in the JSON
Search queries generated by the model	content.text of messages whose recipient is web.run
Candidates returned by search	url / title / snippet in metadata.search_result_groups[].entries[]
URLs cited in the answer	metadata.content_references[] of the final message

Conditions:

• Logged in, desktop web, model gpt-5-6
• Memory off, every custom instruction field empty
• New chat each time; JSON captured only after the answer finished
• Five prompts about SEO and GEO tools, run on two days

First, what the logs don't contain. The exact string the model received is not recorded. The JSON snippet is a display value truncated with …. When the model fetches a page's body, that body isn't logged either. Everything observed here is limited to what the client side records.

ChatGPT cuts the first 200 characters of body text

Across the 383 snippets, the minimum length was 200 characters, the median 202 and the maximum 206. 100% fell within 200–206 characters, and 100% were truncated at the end. The cut is made mechanically by character count, regardless of sentence boundaries.

The content makes it obvious these aren't deliberately written descriptions.

Found in the snippet	Share
Runs of capital letters (traces of English headings rendered in uppercase)	78%
Bullets starting with * (global navigation)	20%
Image: (image alt attributes)	17%
[Input] [Button] [Select] (form elements)	14%

A meta description never contains navigation or form elements. This is text taken from the rendered page.

But it isn't the first 200 characters of the raw page

We tested this directly. We took the first 24 characters of each snippet, searched for them in the raw text of the same page fetched ourselves, and measured the offset where the snippet began.

Of the 12 pages we could match, 0 started at the top (character 0–5) and 12 had their opening discarded.

The discarded openings were all of the same kind:

• Skip to content / Skip to main content
• Rows of global navigation links
• Browser warnings and campaign banners

For example, 1,506 characters of boilerplate were discarded on Semrush's Japanese page, 1,036 on Screaming Frog, and 3,364 on TechRadar; the snippet started in the body text that followed.

In other words, body extraction (boilerplate removal) happens first, and the first 200 characters of the result are cut.

title is recorded as a separate field, and it matched the start of the snippet in only 13% of cases. The title and the snippet are different things.

The meta description wasn't used

The deciding evidence was a one-word difference.

Screaming Frog's product page lists supported operating systems differently in the meta description and in the body.

Figure

Comparing three places on the same page

meta description

In the HTML head

• The industry leading website crawler for Windows, macOS and Ubuntu, ...
• Did not appear in the snippet

Page body (right under the h1)

Rendered text

• The industry leading website crawler for Windows, macOS and Linux, ...
• The snippet matched this one

The snippet recorded in the conversation log was:

SCREAMING FROG SEO SPIDER The industry leading website crawler for Windows, macOS and Linux, trusted by thousands of SEOs and agencies worldwide for technical SEO site audits. * Overview * User Gui…

It contains the body's Linux, not the meta description's Ubuntu. The leading SCREAMING FROG SEO SPIDER is the h1 rendered in uppercase.

We fetched and compared the meta descriptions of five more pages (Rakko Inc., Nobilista, GMO AI Kantan Shukyaku, Ahrefs Webmaster Tools, and Nobilista's pricing page). None of them had meta description wording in the snippet.

Nobilista's homepage, for instance, has a 145-character meta description, but its snippet contained the global navigation. GMO AI Kantan Shukyaku's meta description is only 14 characters, and its snippet began with the h1 "集客を「かんたん」に最大化する" ("Maximize customer acquisition, the easy way").

This differs from Google

Google's official documentation states that it builds snippets from page content and the meta description. Material Google uses was not being passed to ChatGPT.

We want to be careful here. We only tested ChatGPT. AI Overviews and Gemini, which build on Google Search's index, may well make use of the meta description. We have not tested them.

The point isn't that "meta descriptions are obsolete." It's that they don't count as ChatGPT optimization.

Changing the query doesn't change the snippet

On the same day, 15 URLs were picked up by different questions and different search queries. For all 15, the snippets matched exactly.

In Google's results, the excerpt shown for the same page changes with the query. This ChatGPT field doesn't. It isn't looking at the query.

The flip side is that there's no room to tailor snippets per query. There is only one lever: what you put in the first 200 characters.

Body extraction can fail

What happens across days? Of the 89 snippets that appeared on both days, 73 (82%) matched exactly and 16 changed.

The query was not the cause.

Cause of change	Count
Relative date display only ("Updated today" → "Updated yesterday")	2
Body extraction started at a different point	8
The page content itself changed	6

A changing relative date also means snippets aren't a fixed cache; they're rebuilt on every fetch. And each rebuild can produce a different extraction result.

The most common cause was the shifted extraction point. Same page, same 200-character slot:

Figure

Same page, one day apart: the body gets through, or it doesn't

Sep 5: started at the heading

Reached the opening of the body

• PEEC AI VS SEMRUSH AI VISIBILITY TOOLKIT
• TOMEK RUDZKI GEO Expert Jan 30, 2026
• If you're evaluating tools to monitor your…

Sep 6: started at the navigation

Did not reach a single character of the body

• Not sure which is right for you? Talk to us
• Product Pricing Resources Partnerships Careers Log in Sign up
• (110 characters used up by this point)

On Sep 6, navigation and CTAs alone used up more than half the slot. Boilerplate that had been removed the day before was not removed that day. The same page swung between "body gets through" and "body doesn't," depending on the day.

There were other extraction failures. Nobilista's homepage snippet was:

クラウド型SEO対策検索順位チェックツール * ご利用プラン（料金） * 計測プラン * 機能 * 使い方 * お知らせ * 資料請求 * ブログ * [Select] * ログイン * お問い合わせ * 0円で始める クラウド型SEO対策検索順位チェックツール 今ならNOBILISTAの全機能が 無料で使える７日間の無料トライアル 実施中！ 今すぐ0円で利用開始する ※クレ…

Gloss: "Cloud-based SEO rank-tracking tool" followed by the navigation items (Plans (pricing), Tracking plans, Features, How to use, News, Request materials, Blog, [Select], Log in, Contact, Start for 0 yen), then "Cloud-based SEO rank-tracking tool. All NOBILISTA features free with a 7-day free trial, running now! Start for 0 yen now. ※Cre…"

Nearly half of the first part is global navigation. What filled the rest of the slot was a campaign announcement, and not one character explains what the product actually does.

What the failed pages had in common

Both pages where extraction failed (Nobilista's homepage and Peec AI's comparison page) had neither main nor article.

Page	Extraction	main	article
co.nobilista.com	Failed	No	No
peec.ai (comparison page)	Failed	No	No
screamingfrog.co.uk	Succeeded	Yes	No
rakko.inc	Succeeded	Yes	No
kantan.gmo	Succeeded	Yes	No
ahrefs.com	Succeeded	No	No

This is an observation of six pages, so it doesn't show that "adding main guarantees success." Ahrefs succeeded without main. But the fact we did see is that both failed pages declared no body region. That's reason enough to give the extractor more to work with.

Figure

Protect the first 200 characters of your body text

What to do

• ✓Wrap the body in main or article to declare where the content is
• ✓Put a one-sentence description of your company right under the h1
• ✓Use header / nav / footer elements for the header, navigation and footer
• ✓Actually extract and read the first 200 characters of your own body text

What to skip

• ✕Count meta description polishing as ChatGPT optimization
• ✕Stuff keywords into the first 200 characters
• ✕Try to serve different snippets per query
• ✕Fill the opening of the body with nothing but campaign announcements

Snippets get paraphrased into the backbone of the introduction

So how do those 200 characters show up in the answer? Here are real examples side by side.

Nearly verbatim (Ahrefs)

Snippet (ahrefs.com/ja/seo-audit-tool)

無料版 SEO ツール / 無料の SEO 監査ツール **170 以上の SEO に関する一般的な問題を見つけて修正しましょう。**インストール不要、SEO の高度な知識も不要です。[Input: ドメインを入力] 無料で登録 …

Gloss: "Free SEO tools / Free SEO audit tool. Find and fix 170+ common SEO issues. No installation, no advanced SEO knowledge required. [Input: enter your domain] Sign up free …"

Answer

無料版でもSite Auditは1プロジェクトあたり月5,000クロールまで利用でき、170種類以上のSEO問題をチェックできます

Gloss: "Even on the free plan, Site Audit allows up to 5,000 crawls per project per month and can check for 170+ types of SEO issues."

Only the meaning survived (Rakko Keyword)

Snippet (rakko.inc/services/keyword/)

ラッコキーワード Image: ラッコキーワードのロゴ 膨大な検索データを効果的に分析するマーケティングツール インターネット上の膨大な検索データ（キーワードデータ、SERPデータ）を効果的に分析できる形で集約し、ユーザーニーズの抽出に大きく役立つマーケティングツールです。

Gloss: "Rakko Keyword. Image: Rakko Keyword logo. A marketing tool for effectively analyzing vast search data. A marketing tool that aggregates the vast search data on the internet (keyword data, SERP data) into an analyzable form and greatly helps extract user needs."

Answer

個人でコンテンツSEOをするなら特に使いやすい候補です。関連キーワードや検索ニーズの調査などに強く

Gloss: "An especially easy-to-use option if you do content SEO on your own. Strong at things like researching related keywords and search needs"

The headline became the pitch (GMO)

Snippet (internet.gmo/news/article/119/)

GMO NIKKO、知識がなくても使える Web集客支援ツール「GMO AIかんたん集客」月額2,992円～提供開始

Gloss: "GMO NIKKO launches GMO AI Kantan Shukyaku, a web customer-acquisition tool usable without specialist knowledge, from 2,992 yen/month"

Answer

SEOの知識があまりない個人事業主なら有力です

Gloss: "A strong option for sole proprietors without much SEO knowledge"

Copy the companies wrote themselves comes out of the AI's mouth, paraphrased.

It doesn't necessarily come out verbatim

We measured overlap between the answer text and the set of snippets using runs of 8 consecutive matching characters. It was 6–11%. 90% of the answer does not match the snippets verbatim.

That's why keyword stuffing doesn't work. What carried through, as in the three examples above, was a single sentence shaped as a claim. The sentence "usable without specialist knowledge" survived with its meaning intact, even after being paraphrased.

But figures come from the body

Tidying up the first 200 characters isn't the whole job.

The answer about GMO AI Kantan Shukyaku includes the figure "3,840 yen for a single month (excl. tax)." That string appeared nowhere in the 383 snippets. When we fetched the press release page shown as the citation, the figure was there in the body, verbatim.

The same answer also says "4,224 yen including tax," which appears in none of the sources. It's 1.1 × 3,840 yen: a figure the model calculated and added on its own. The arithmetic is correct, but a number not stated in the source appears in a cited context.

Figure

The first 200 characters and the body play different roles

Step 1

First 200 characters

Material the page is judged on in the candidate list

If it fails here: If navigation fills it, no product description gets through

↓→

Step 2

How you're introduced

Paraphrased into the backbone of the introduction

If it fails here: Sentences not shaped as claims don't survive

↓→

Step 3

Page body

Source of figures such as pricing and specs

If it fails here: Figures not in the body go unmentioned, or get calculated

↓→

Step 4

Answer

Introduction plus figures, assembled

A previous article on this site covered a case where no tool call fetching a page body was recorded at all, and every fact in the answer could be explained by strings inside the snippets. This time that wasn't the case. Both happen. The first 200 characters alone aren't enough, and neither is the body alone.

Stable snippets don't mean stable answers

This is easy to misread, so let's be explicit.

Over the same two days, 73 of 89 snippets matched exactly. But the answers changed.

For the question about GEO tools, Profound was listed first on Sep 5, but on Sep 6 Otterly came first, with the flat recommendation "if you pick just one, OtterlyAI." Peec AI, listed on Sep 5, disappeared from the answer and Scrunch took its place.

In a Japanese question with three conditions attached, one of the three slots also switched to a different product. On Rakko Keyword's pricing, Sep 5 said "from 1,320 yen on monthly billing," while Sep 6 said "Entry plan at 660 yen on monthly billing." 660 yen is the official price on annual billing, so the Sep 6 statement is most likely inaccurate.

Same input, moving output. Tidying the first 200 characters is work to get you "introduced correctly once you're a candidate," not work that locks in a ranking. Don't judge from a single observation; measure several times, spaced out over time.

What this observation shows

• All 383 snippets recorded in the conversation JSON were 200–206 characters and truncated at the end
• On all 12 pages we could match, the header, navigation and skip links had been discarded (0 started at the top of the page)
• On the 6 pages checked, no meta description wording appeared
• The 15 URLs picked up by different queries on the same day had exactly matching snippets
• Across days, 82% matched, and the main cause of change was a shift in the body extraction point
• Both pages where extraction failed had neither main nor article

The sample is 5 prompts × 2 days, and we didn't look at anything other than ChatGPT. Since Google states that it uses the meta description, Google's products could show the opposite result.

How to check your own pages

1. Turn off memory, clear custom instructions, and open a new chat
2. Ask a question in your category and capture the conversation JSON after the answer has finished
3. Confirm there's no model_editable_context message (if there is, memory is active)
4. Find your domain's snippet in search_result_groups[].entries[]
5. Compare it against your page's meta description. It probably won't match
6. Look at what's at the start of the snippet. If it's navigation, body extraction failed

If you enter a URL into our page audit tool, it shows the first 200 characters of body text with the header, navigation and footer removed, as is. For pages that don't declare a body region, it flags the risk that extraction will be replaced by navigation.

Note that the audit reads the HTML the server returns, and we haven't confirmed whether ChatGPT sees the page after JavaScript execution. The extraction algorithm isn't the same either. Use it as a guide.

FAQ

Does that mean I can stop writing meta descriptions?

No. Google officially states that it builds snippets from page content and the meta description, so it matters for how you appear in Google Search. The only thing this observation shows is that meta description wording didn't appear in the snippets recorded in ChatGPT's conversation logs. The takeaway is to stop counting meta description work as ChatGPT optimization.

Is it the same for Gemini or AI Overviews?

We don't know; we haven't tested them. If anything, since AI Overviews builds on Google Search's index, the meta description is likely to be used. Don't extend conclusions about ChatGPT to AI search in general.

What should I put in the first 200 characters?

One sentence stating what your service does. The sentences that survived into answers this time were shaped as claims, such as "usable without specialist knowledge," "greatly helps extract user needs," and "find and fix 170+ common SEO issues." Keyword lists didn't survive. Verbatim overlap between answers and snippets was 6–11%, so the text doesn't come out as-is.

How do I keep navigation out of the snippet?

Normally it's removed automatically: on all 12 pages we could match, the header and skip links had been discarded. The problem is when removal fails, and both pages that failed this time declared no body region. What you can do now is wrap the body in main or article and mark up the header, navigation and footer as header / nav / footer elements. Note that data-nosnippet is an attribute for controlling Google's snippets, a separate mechanism from what we observed. We haven't confirmed whether ChatGPT respects it.

Are pages with empty snippets at a disadvantage?

Not necessarily. This time, the 24 citations shown in the most prominent form within answers (a link right under a heading) all had empty snippets. Pages with snippets ended up as smaller citations placed alongside. What the presence or absence of a snippet means hasn't been disclosed, and it shouldn't be read as better or worse.

Is checking once enough?

No. This time, 82% of snippets matched across the two days, yet the order of tools in the answer shifted, one dropped out and another came in. In one case, the quoted price changed for the same question. Measure several times, spaced out over time, and record the results as ratios.