Skip to content
Get Ranked On TopIllustrative case studyContent Engine plan

A fictional business. The figures are a modelled scenario, not a client result.

XERBA LABS · Arabic speech-to-text API, 18 staff · Tunis

How an Arabic speech-to-text startup got AI assistants to quote its own docs

+348%

API sign-ups from search and AI answers

Tunis, Tunisia · 12 months · Technology & SaaS · Content Engine plan

The stone grinds, the oil runs: every drop is a passage an AI assistant can quote.

01THE CHALLENGE

The answers were pressed from someone else’s olives.

Developers ask AI assistants which Arabic speech-to-text API to use, and the answers cited big-cloud docs. Xerba's thin docs were never quoted.

Developers no longer start on a results page. They ask an AI assistant which Arabic speech-to-text API to use, paste the answer into a terminal and move on. When we ran 140 such prompts, ten times each, the answers leaned on the documentation of the large cloud providers, most of which cover Modern Standard Arabic and little else.

Xerba’s models handle Tunisian, Moroccan and five other varieties better than anything those answers recommended. The docs never said so in a form a model could lift: no definitions, headings named after fields, samples that failed on paste, and accuracy claims without a single number.

Baseline: Xerba was cited in 3 of 140 tracked prompts.

WHAT THE HARVEST AUDIT FOUND

  1. 0 / 61

    No definitions

    None of the 61 docs pages defined the term it was about. “Dialect detection” and “diarization” appeared everywhere and were explained nowhere.

  2. 73%

    Headings named after fields

    73% of headings were nouns such as “Parameters” or “Overview”. Developers ask in tasks: transcribe a call, split two speakers.

  3. 4 / 38

    Samples that didn’t run

    Only 4 of 38 code samples ran as written. The rest used a retired auth header, skipped imports or ended in an ellipsis.

  4. 1 / 7

    One dialect on paper

    The product handles seven Arabic varieties; the docs named one. With no benchmark table, assistants had nothing to compare and fell back on big-cloud pages.

02THE STRATEGY

Write docs a model can press.

Built on our Content Engine plan: content and AI-search work for teams whose buyers ask machines before they ask people.

  1. 01DEFINITION-FIRST DOCS

    61 pages rewritten

    Every page opens with a definition

    A 40–60-word answer to “what is…?” right under the title, naming the term, the endpoint and its limits. It reads well to a skimming human and survives being cut out of the page whole.

  2. 02TASK HEADINGS

    212 task headings

    Headings phrased as tasks

    “Parameters” became “Transcribe a Tunisian call recording”. Each heading matches a question developers actually type, and the paragraph under it answers only that question.

  3. 03RUNNABLE SAMPLES

    38 / 38 samples pass

    Samples that run on paste

    Every snippet in curl, Python and Node runs against the live API each night; a failing sample blocks the docs deploy. The expected output sits right under the code.

  4. 04COMPARISON TABLES

    7 dialect benchmarks

    Tables with real numbers

    Word error rate by dialect, latency by file length and an honest comparison page with its method in full. Assistants quote tables almost verbatim, so every number had to be defensible.

03THE EXTRACTABILITY PRESS

Put one docs page through the press.

AI assistants don’t cite pages, they cite passages: a definition that stands on its own, a heading that names the task, code that runs, a table with numbers. Switch on each fix and watch which passages the press can squeeze out as citations.

Fixes to apply
xerba.example/docs/transcribeold version
  1. Overview

    Xerba provides powerful, innovative speech solutions for the Arabic-speaking world.

    1 · Extractability 14/100Left in the pulp

    A slogan, not a definition: nothing a model can quote as fact.

  2. Parameters

    audio, dialect, diarize, callback_url. See the API reference for details.

    2 · Extractability 22/100Left in the pulp

    “Parameters” matches no question anyone asks.

  3. Quick start

    Initialise the client with your key, then call transcribe with your options.

    client = Xerba(key)
    result = client.transcribe(file, ...)
    3 · Extractability 9/100Left in the pulp

    Placeholder code: an assistant won’t recommend what it can’t verify.

  4. Accuracy

    Xerba supports many Arabic dialects with industry-leading accuracy.

    4 · Extractability 17/100Left in the pulp

    “Industry-leading” is an adjective, not evidence.

In this model, assistants quote passages that score 70 or more.

Passages pressed out

0 of 4

What an assistant answers

Developer prompt

which api should i use for tunisian arabic speech to text?

General-purpose cloud speech APIs mostly support Modern Standard Arabic; dialect coverage varies by provider.[1]

None of the sources describe a dialect-specific option, so a general cloud API is the safe suggestion.

Sources

  1. 1.General cloud speech-to-text docs

0 of 4 passages pressed out. Xerba not cited.

04MONTH BY MONTH

Twelve months, from harvest to bottle.

  1. MONTHS 1–201

    Harvest

    Audit

    Crawled all 61 docs pages, ran 140 developer prompts ten times each across four assistants, and logged every source they quoted.

    3 of 140 prompts cite Xerba

  2. MONTH 302

    Wash

    Structure

    Merged duplicate guides, gave every page one job and one URL, and wrote the glossary the rest of the docs link to.

    4 of 140 prompts cite Xerba

  3. MONTHS 4–603

    Grind

    Rewrite

    Definition-first intros and task headings for all 61 pages, Tunisian and Moroccan guides first because nobody else had written them.

    14 of 140 prompts cite Xerba

  4. MONTHS 7–804

    Press

    Samples and tables

    Nightly-tested samples in three languages, dialect benchmark tables and a comparison page with its method in full.

    25 of 140 prompts cite Xerba

  5. MONTHS 9–1005

    Decant

    Measure

    Monthly prompt runs showed which passages were quoted and which were only paraphrased; weak passages were rewritten, not padded.

    36 of 140 prompts cite Xerba

  6. MONTHS 11–1206

    Bottle

    Compound

    New guides for call centres and media archives, French versions of the top 20 pages, and a public changelog assistants now cite.

    46 of 140 prompts cite Xerba

05RESULTS

A year on, the answers pour from Xerba’s own docs.

3 → 46 of 140

Tracked AI prompts citing Xerba
46 of 1403M1M6M12

+348%

API sign-ups from search and AI answers
18842M1M6M12

1,900 → 11,400

Organic docs clicks / month
11,4001,900M1M6M12

19 → 6 min

Time to first transcript
Before19 minAfter6 min

Modelled scenario for a fictional business — not a client result. Rankings, AI citations and traffic are never guaranteed; see our Terms.

SIX OF THE 140 TRACKED PROMPTS

How often Xerba was cited in ten runs of each prompt, month 0 and month 12.
Prompt, as typedMonth 0Month 12Share of runs
best api for tunisian arabic speech to text0 / 109 / 10
arabic speech recognition api with dialect detection1 / 108 / 10
how to transcribe moroccan darija audio in python0 / 107 / 10
arabic stt api word error rate comparison0 / 106 / 10
API transcription arabe dialecte tunisien0 / 109 / 10
speaker diarization arabic call center api0 / 105 / 10

A prompt counts as citing Xerba when at least five of its ten runs quote the docs.

YOUR DOCS, PRESSED

Get a plan like this

If developers ask an AI assistant before they open your docs, start with the Content Engine plan, or tell us what your API does.

← All case studies