← cd ../case-studiesGet Ranked On Top · SEO case study · Asia-Pacific
Last login: modelled session on ttys011 · Singapore
mt case-study merlintide
clientMerlintide
sectorDocument OCR API for developers · Singapore
industryTechnology & SaaS · 11 months
How an OCR API won developer searches with tutorials engineers actually bookmark
# headline
+305% API key sign-ups from organic
sparkline · M00 → M11
M00 █████··············· 100 M01 █████··············· 105 M02 ██████·············· 117 M03 ███████············· 133 M04 ████████············ 155 M05 █████████··········· 180 M06 ██████████·········· 209 M07 ████████████········ 241 M08 ██████████████······ 277 M09 ████████████████···· 317 M10 ██████████████████·· 359 M11 ████████████████████ 405
cat CHALLENGE.md
The challenge
Merlintide's API was faster and cheaper than rivals, but developers searching 'extract table from PDF python' only found hyperscaler docs.
The product had already won on paper: lower latency, a cheaper per-page price and cleaner JSON. But engineers don't search for “OCR vendor”. They search for the task in front of them, in the language they're writing, usually mid-sprint. Merlintide had no page that answered a single one of those tasks, and its docs were invisible to crawlers.
Meanwhile, more of those questions were being asked to AI assistants first, and the assistants quoted whoever had the clearest runnable answer. That was never Merlintide.
# query → who answered (M00) "extract table from pdf python" → hyperscaler docs #1 → Q&A forum #2 → merlintide.example #41 # AI assistant citations: 0
mt audit docs.merlintide.example
Diagnosis
A crawl of the docs, the logs and 312 developer queries, scored the way a search engine and an AI assistant would see them. Seven findings, three blocking.
- [FAIL] 01
render: Docs were a client-rendered SPA: 0 of 214 code samples present in the initial HTML.
fix: Static pre-render of every docs and tutorial route; code blocks shipped as real text.
- [FAIL] 02
content: No pages for the 40 highest-intent task queries ("extract table from pdf python", "ocr invoice node").
fix: Task-first tutorial library mapped one-to-one to those queries.
- [FAIL] 03
schema: No TechArticle, HowTo or SoftwareSourceCode markup anywhere on the docs.
fix: Code-sample schema on every tutorial, linked to the API reference entity.
- [WARN] 04
dupes: /v1, /v2 and /latest reference pages indexed side by side with no canonical.
fix: One canonical per endpoint; old versions noindexed and cross-linked.
- [WARN] 05
answers: Tutorials opened with marketing copy; nothing an AI assistant could quote in one pass.
fix: 40-word answer block plus a runnable snippet as the first thing on every page.
- [WARN] 06
speed: Docs LCP 4.1 s on mobile (syntax highlighter blocking the main thread).
fix: Highlighting moved to build time; LCP modelled at 1.3 s.
- [PASS] 07
product: Median latency and per-page price beat the hyperscaler APIs in public benchmarks.
fix: Benchmark page with a reproducible notebook became the link-earning asset.
3 failed, 3 warnings, 1 passed in 4.2s
cat plan.yml
The strategy
plan: Content Engine duration: 11 # months goal: "be the answer engineers paste, then bookmark" tactics:
- name: "Developer tutorials"
flag: --tutorials=48
Forty-eight task-shaped tutorials (one query, one outcome) in Python, Node, Go and cURL, each ending in a free-key snippet that runs on the first paste.
- name: "Code-sample schema"
flag: --schema=SoftwareSourceCode
TechArticle + SoftwareSourceCode markup with language, runtime and dependency fields, so search engines can read a sample as code rather than prose.
- name: "AI Overviews"
flag: --answers=first
Every page leads with a quotable answer, a minimal code block and a version-stamped date: the format AI assistants cite when engineers ask them how.
- name: "Docs SEO"
flag: --docs=static
Pre-rendered docs, one canonical per endpoint, a crawlable sidebar and internal links from each tutorial into the exact reference page it uses.
# the Content Engine plan sets the cadence; the tactics above are how it was pointed at developer intent.
git log --reverse --oneline
11 months, commit by commit
- a41c9e2M01105
audit: crawl docs, map 312 developer queries
Search console, log files and competitor docs crawled; intent grouped by task.
- 7be03d1M02117
feat(docs): static pre-render + canonicals
Code samples visible to crawlers for the first time.
- c90f4aaM03133
feat(tutorials): first 8 python tasks
Tables, invoices, receipts, handwriting; each with a runnable snippet.
- 13d7e6bM04155
feat(schema): SoftwareSourceCode on samples
Validated against the rich result tests; zero errors.
- e2a58f0M05180
feat(tutorials): node + go ports
Same tasks, three more runtimes, hreflang-style language switcher.
- 5f61b3cM06209
perf: build-time highlighting, LCP 1.3s
Main-thread work on docs pages cut by roughly two thirds.
- 98c2d47M07241
feat(answers): answer-first rewrite
First AI assistant citations of tutorial pages logged.
- 0ad4e19M08277
feat(bench): reproducible benchmark notebook
Picked up by two newsletters and a framework's docs.
- b7e3f82M09317
feat(tutorials): 48 tasks live
Long-tail coverage complete; internal links into reference.
- 4c18ad5M10359
chore: refresh samples for SDK v3
Version stamps updated so answers stay quotable.
- f03e6c9M11405
release: sign-up flow inside tutorials
Free key issued inline; organic sign-ups at their modelled peak.
# right column: organic API sign-ups, index M00 = 100
curl -s mt.dev/cs | jq .metrics
Results
{
"api_sign_ups_from_organic": "+305%",
"tutorial_sessions": "+460%",
"ai_assistant_citations": "86",
"top_3_developer_queries": "+148",
"months": 11,
"modelled": true
}Modelled scenario for a fictional business — not a client result. Rankings and traffic are never guaranteed; see our Terms.
$ mt chart api_sign_ups_from_organic
+305%
API sign-ups from organic
M00 100M11 405
$ mt chart tutorial_sessions
+460%
Tutorial sessions
M00 100M11 560
$ mt chart ai_assistant_citations
86
AI assistant citations
M00 0M11 86
$ mt chart top_3_developer_queries
+148
Top-3 developer queries
M00 0M11 148
mt rank --top 10
Rankings & the SERP, before and after
# 10 of 10 queries · press / to focus
| query | intent | vol/mo | M00 | M11 |
|---|---|---|---|---|
| "extract table from pdf python" | task | 6,600 | #41 | #2 |
| "ocr api" | commercial | 4,400 | #27 | #3 |
| "invoice ocr node js" | task | 1,300 | — | #1 |
| "receipt ocr api free tier" | commercial | 880 | — | #2 |
| "pdf to json python" | task | 3,600 | #63 | #3 |
| "handwriting recognition api" | commercial | 2,400 | #38 | #4 |
| "parse scanned pdf golang" | task | 480 | — | #1 |
| "tesseract vs cloud ocr" | compare | 1,900 | #52 | #2 |
| "ocr bounding boxes json" | task | 590 | — | #2 |
| "extract text from image curl" | task | 1,000 | #74 | #5 |
- #1cloud-hyperscaler.example/docsDocument AI quickstart
- #2qa-forum.exampleHow to read a PDF table? [closed]
- #3other-cloud.example/docsTextract developer guide
- -#41merlintide.exampleMerlintide | Fast OCR API
- +AIAI overviewcites merlintide.example/tutorials/pdf-tables
- #1cloud-hyperscaler.example/docsDocument AI quickstart
- +#2merlintide.example/tutorialsExtract tables from a PDF in Python (9 lines)
- #3qa-forum.exampleHow to read a PDF table? [closed]
mt plan --like merlintide
Get a plan like this
Built on Content Engine: tutorials and docs shaped around the tasks your buyers actually type, marked up so search engines and AI assistants can read the code.
$ cd ../case-studies # all case studies