What a search result costs an LLM: HTML vs JSON vs markdown, measured on 6 AI surfaces

Key findings

Tokens per answer, by format and surface

Each cell is the median over the prompts that surface answered. "JSON" is cloro's full response object, compact, without the markdown and HTML fields. For AI Overview it is the AI Overview block alone, not the whole results page. "Plain text" is the answer with no links, the floor for any format that keeps the words.

The takeaway: the format choice that matters most is HTML against anything else. Every structured format lands between about 450 and 4,900 tokens per answer. Raw HTML is 150 to 750 times larger than the markdown of the same answer. Markdown against JSON is the smaller, second decision.

Raw HTML breaks the budget, and often the context window

Most of an AI answer page is code the answer does not need. On Gemini, scripts, style sheets and SVG icons make up about 95% of the page's tokens. We measured the HTML exactly as the page rendered for a logged-out session, the same file a do-it-yourself browser scraper would capture. OpenAI prices input above 272K tokens at the long-context rate, which for gpt-6.1-terra is $2.00 per 1M instead of $1.00 (OpenAI pricing, read 2026-11-10). The model's context window is 1,050,000 tokens with a 922,000-token input cap (model page).

Gemini, both AI Mode views and every AI Overview results page went over 272K on all 20 prompts. ChatGPT went over on 9 of 20, and three ChatGPT pages ran past 1.1 million tokens, more than the model can take at all. Bing AI Mode was the lightest at a median 207,781, under the threshold on every answer.

With every <script>, <style>, <svg> and comment removed, the median page still ran from 28,551 tokens on Gemini to 163,502 on Bing AI Mode. Across the six surfaces, that is 22 to 142 times the median markdown of the same surface, with AI Overview the highest. What remains is markup and interface text, and the model still has to find the answer inside it. If you scrape AI answers yourself, budget for the parsing step that turns the page into text; without it, every model call pays for the whole page.

Markdown cuts full JSON by 31% on ChatGPT and by 81% on AI Overview

cloro's API returns each LLM answer as JSON: the answer text plus structured lists of the sources and of every inline citation, each with its label, URL, description snippet, position and redirect link. The markdown keeps the answer and writes each citation as an inline link. So the saving tracks how much citation detail a surface returns.

AI Overview saves the most because its JSON block carries long citation entries for every source chip. ChatGPT saves the least, 31%, with a range of 2% to 71% across the 20 prompts. Its answers are longer and carry fewer inline citations, so the JSON has less to cut. Gemini shows the same effect at the edge: on 3 of 20 prompts it answered without any sources, the JSON was just the text, and the markdown came out the same size or up to 7% larger because of its formatting.

Pretty-printed JSON makes it worse: indentation added a median 8% to 9% on most surfaces, 18% on Bing AI Mode and 26% on ChatGPT. If you send JSON to a model, send it compact.

If you already trim your JSON, markdown saves 10% to 18%

Many pipelines do not send the full response. They keep the answer text and the source list and drop the rest. We measured that trimmed JSON too. Against it, markdown's lead shrinks to a few hundred tokens on the chat surfaces.

Most of markdown's saving over full JSON is citation detail a pipeline would drop anyway. What markdown adds is convenience: each citation already sits next to the claim it supports, and you write no trimming code. AI Overview is the exception, where markdown still saves 43% over the trimmed JSON because its source entries are long.

What 1,000 answers cost at OpenAI's prices

Input cost per 1,000 answers at gpt-6.1-terra's $1.00 per 1M input tokens, the long-context $2.00 rate applied to each HTML answer above 272K. At gpt-6-luna ($0.10) divide by ten; at gpt-6.1-sol ($2.00) double it.

At a few thousand answers a day the gap between JSON and markdown is a few dollars a month. The gap between HTML and either one is far larger: a pipeline that reads 10,000 AI Overview pages a day as raw HTML spends about $14,300 a day on input, against about $10 as markdown.

AI Mode on mobile returns the same lean markdown

We ran AI Mode twice per prompt, once as desktop and once as mobile, since the two layouts differ. The mobile markdown was a median 924tokens against 1,044 on desktop, and saved 65% against JSON. We read all 20 mobile markdown files by eye and found no interface text, HTML tags or broken links. Mobile answers carried slightly fewer inline links, a median of 7 against 9 on desktop, with the same median of two sources.

How SerpApi's markdown figures compare, and why they measure something else

SerpApi added markdown output in August 2026 and reports that it gives "roughly half the tokens of JSON on average" across its APIs, from 36% on Walmart to 91% on Apple Maps (SerpApi, August 21, 2026). In a later test, also counted with o200k_base, SerpApi measured one Google search for "coffee" at 66,237 tokens as JSON and 13,025 as markdown, and compared it with Exa, Tavily and Firecrawl, whose full-page modes ran 30,802 to 133,343 tokens (SerpApi, October 1, 2026).

Those figures are about organic Google results: the ten blue links and the features around them. This study measures AI answers. cloro returns organic Google results as JSON and HTML only, with no markdown option, so there is no cloro number to set beside SerpApi's. The queries, locations and payloads differ too, so the two sets of numbers cannot be compared directly.

Which format to send: markdown for models, JSON for data, HTML for neither

• For an agent or a RAG pipeline, send markdown. It is the smallest format that keeps the citations next to the claims they support, and it needs no clean-up code.
• For dashboards and citation tracking, keep the JSON. Source positions, domains and citation entries are fields you count, and markdown flattens them into text.
• To do both, request both formats in one call. Store the JSON, send the markdown to the model.
• Keep raw HTML for audits and screenshots. At 200,000 to 700,000 tokens a page it is the wrong input for a model.

With cloro the flow is one request: send the prompt with include.markdown set to true, cloro runs the query and parses the answer, and you get the JSON and the markdown back together. For AI Overview, request include.aioverview with markdown: true. Markdown costs no extra credits (provider docs).

curl -X POST https://api.cloro.dev/v1/async/task \
  -H "Authorization: Bearer $CLORO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "taskType": "AIMODE",
    "payload": {
      "prompt": "Which provider gives the most reliable Google AI Mode coverage?",
      "country": "US",
      "include": { "markdown": true }
    }
  }'

If you are moving from another provider, the migration guide maps SerpApi, DataForSEO and four other vendors to cloro's fields, and thellms.txt guide covers the same token problem from the publisher's side.

Methodology

We took 20 prompts from the 176 buyer questions cloro tracks in AI engines (every ninth one, covering SERP API buying, AI visibility tracking and AI scraping) and sent each one to ChatGPT, Gemini, Google AI Mode on desktop and on mobile, Bing AI Mode, and Google Search with the AI Overview, through cloro's async API, country US, on November 10, 2026. For AI Overview the prompt was the Google query, and all 20 returned an AI Overview. Each request asked for markdown and HTML in the same call. One Bing AI Mode request failed twice and is excluded, so Bing has 19 answers. Perplexity was not measured: its API was unavailable on the day.

Formats: raw HTML is the rendered page snapshot cloro stores for each request. Stripped HTML removes script, style,svg, noscript, template, link and meta elements and comments. JSON is the result object serialized compact, without the markdown and HTML fields; for AI Overview, the AI Overview object alone. Trimmed JSON keeps only the answer text and the sources list. Markdown is the response's markdown field. Plain text is the answer text field.

Tokens are counted with tiktoken's o200k_base encoding, used by OpenAI's gpt-4o generation. OpenAI does not publish the encoding of its newer models, so the counts are a proxy for them. A second count withcl100k_base came out a median 1.1% higher on the markdown. Every figure is a median of per-answer values; savings are the median of each answer's own saving. Prices are OpenAI's standard input rates as published on November 10, 2026. Answers vary by day and by session, so treat these as a one-day measurement on one prompt set.

Data: the token counts for every answer and format, with the prompt text, are free to reuse under CC BY 4.0:json-vs-markdown-tokens-2026-11-10.csv.

More studies from cloro's monitoring corpus are in the research index.

Questions this study answers