iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
When SerpApi returned a Google search for “coffee” as full JSON, the response measured 24,723 tokens. The same search returned as full Markdown measured 6,435 tokens, a 74% reduction in SerpApi’s own example. Restricting the response to organic results cut the JSON to 8,486 tokens and the Markdown to 1,298.
The counts are measured. The explanation of which parts of the response cost the most is not. SerpApi has not published the token cost of each field, so this article separates the figures you can rely on from the reasoning about where the overhead likely sits.
The numbers, side by side
SerpApi published four counts for one Google search for “coffee.” They form a grid: two output formats, each returned in full or restricted to organic results.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Response for the “coffee” Google search | Full response | Organic results only |
|---|---|---|
| JSON | 24,723 tokens | 8,486 tokens |
| Markdown | 6,435 tokens | 1,298 tokens |
These are SerpApi’s 2026 figures for one query. Restricting JSON to organic results cuts it by about 66% (calculated from the published counts). Markdown alone cuts the full response by 74%. Restricted Markdown is 95% below full JSON, and about 80% below full Markdown (calculated from the same counts). Token totals depend on the tokenizer of the model that reads the text, so your absolute numbers will differ, but the ratios show how much each lever contributes.
#1 Best Overall
Where the overhead likely comes from
MachineLearningMastery.com, in a September 18, 2026 article about the SerpApi feature, lists several features of a JSON search response that probably add tokens without adding much answer content. Treat these as the likely sources, not as measured line items.
Internal tracking links
Each result can carry links that exist for the search engine’s own bookkeeping. A language model reading the response has no use for most of them, but it still has to process every character of each URL.
Icons and thumbnails
Image and icon references are stored as addresses and attributes. They take space in the payload even when the text they accompany is what the reader actually needs.
Free tools Windows power users keep installed
One-click scans. No signup required.
Nested metadata
JSON represents structure through nested objects and keys. Every level of nesting adds punctuation and key names, and those repeat across dozens of results.
Repeated title and link labels
The same title and link appear under more than one key in some responses. The duplication is harmless to a parser but costs tokens when the text is read in full.
None of these elements has a published token count. The article’s explanation is a reasoned account of why the JSON is longer than the information it carries, which is why the restricted and Markdown figures above are more useful than any per-element estimate.
Rank #3
Two different ways to shrink a response
SerpApi separates two operations that are easy to confuse. Markdown output changes how the same sections are written. JSON Restrictor removes sections or fields you do not request. Each changes a different part of the payload, which is why the two can be combined.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →| Lever | What it changes | What it removes | Output a program can still parse as typed data |
|---|---|---|---|
| Markdown output | Representation: tables, Markdown links, and YAML frontmatter replace typed JSON objects and arrays | Nothing by design; SerpApi says most informational content is preserved | No; it is readable text, not a typed schema |
| JSON Restrictor | Selection: the response contains only the fields or sections you choose | Sections and fields you do not request | Yes; it stays JSON |
SerpApi’s announcement, dated August 21, 2026, puts the distinction this way. Tomás Murúa, identified as the author of the post, writes: “The difference with Markdown output is what each one removes.” A second line from the same post makes the contrast sharper: “One subtracts data, the other changes its shape.”
How to request Markdown
SerpApi documents three ways to ask for Markdown in a search request:
- Add
output=mdas a parameter to the search request. - Call the
/search.mdroute instead of the standard search route. - Send the HTTP header
Accept: text/markdown.
The response then contains tables, Markdown links, and YAML frontmatter. To shrink it further, combine the Markdown request with JSON Restrictor, which SerpApi documents as a supported combination. Check the current parameter names in SerpApi’s documentation before you build against them, since the feature was launched recently and its supported APIs may change.
Choosing JSON or Markdown
The right format depends on who reads the response. Use these rules as a starting point.
Recommended Free Tools
- Use Markdown when the next step is a language model reading, summarizing, or reasoning over the results. Its tables and links are easy for a model to scan, and it is far shorter.
- Use JSON when application code depends on exact field names, types, arrays, or predictable parsing. Markdown is text, so a program has to parse it again before it can use the values.
- Use JSON Restrictor with JSON when your code needs only a few fields. You keep the structure and drop what you never read.
- Use restricted Markdown when a model needs only a subset of the results. This combination gives the smallest payload in SerpApi’s example, but it is also the one most dependent on your field selection being correct.
In short, if a language model is the reader, start with Markdown. If your code is the reader, keep JSON and restrict it.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the figures do not establish
The published numbers are vendor measurements of one query. They do not describe every search engine, response, or workflow, and no independent benchmark of them has been published that we could confirm.
- Token count is not total cost. A smaller prompt lowers input tokens. It does not tell you the total cost, latency, or answer quality of your application.
- The average savings figure is a product-page claim. SerpApi’s product page reports average token savings of about 50%, with examples of 74% for Google Search and 90% for Google Shopping. These are SerpApi’s own numbers.
- Feature details can change. SerpApi announced the feature in its weekly changelog on August 18, 2026, and described it in detail on August 21, 2026. Supported APIs, published savings, and availability may have changed since the official pages were reviewed on October 7, 2026.
Measure your own responses
Because the published example is one search, the most reliable figure for your work is the one you measure. A sound test takes about an hour and uses your own queries.
- Pick five to ten queries that match what your application actually sends, including at least one that returns few results and one that returns many.
- Request each query four ways: full JSON, JSON with restriction, full Markdown, and restricted Markdown.
- Count tokens for each response with the tokenizer of the model that will read it.
- Check that every field your code uses still appears in the restricted JSON, and that the Markdown contains the information your model needs to answer.
Record the counts next to the query and the field selection. The ratio between formats will usually matter more than the absolute count, and it will show whether the 74% from SerpApi’s example holds for your searches.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsExplanatory context for the feature comes from MachineLearningMastery.com, a partner-content article dated September 18, 2026 and attributed to its MLM Team. The primary evidence for the feature, the request methods, and the token counts is SerpApi’s own announcement and documentation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
