research · IndicGEO

AI search research in India, with every answer and script published

IndicGEO is Depra AI's open research on how ChatGPT, Gemini and Perplexity answer buyers in India, led by Zia Wasi. Study 01 is published. Every response, prompt and script of its August fieldwork and October re-run is open in the IndicGEO repository on GitHub (opens in a new tab).

Updated

A glass flask holding red, orange and yellow bars, with small glass shapes rising out of it
about the research

The research asks buying questions in English and in Hinglish (Hindi and English typed in English letters). In August 2026, Study 01 found that asking in Hinglish changed which brands Gemini and Perplexity recommend. The October re-run found the gap still there on Perplexity (28.2 points) and down to 2.0 points on Gemini. Zia Wasi is co-founder of Depra AI and a GEO (generative engine optimization) researcher.

Study 0114 Aug 2026, re-run 13 Oct 2026

Ten buying prompts, each asked in English and in Hinglish, run eight times against ChatGPT, Gemini and Perplexity in India. In August 2026, beyond normal rerun noise, the brand list changed 2.2 points on ChatGPT, 7.8 on Gemini and 23.4 on Perplexity. In the October 2026 re-run (479 answers) the gap was 28.2 points on Perplexity, 2.0 on Gemini and 0.8 on ChatGPT.

GeminiIn August 2026 Gemini cited sources on 90.0% of English answers and 41.3% of Hinglish ones; in the October re-run, 73.4% and 56.3%

  • 480 answers (Aug)
  • 479 answers (Oct)
  • ChatGPT, Gemini, Perplexity
  • India-geolocated
  • Data and code public
Depra AI Research · IndicGEO Study 01Both waves

Brand-list gap beyond rerun noise, August and October 2026

Brand-list gap between English and Hinglish beyond rerun noise, by engine, August and October 2026
EngineAugust 2026 gapAugust pOctober 2026 gapOctober p
ChatGPT2.2 pts.0460.8 pts.242
Gemini7.8 pts.00982 pts.063
Perplexity23.4 pts.001028.2 pts.0010
How much more the brand list changed between the English and Hinglish versions of the same question than between two reruns of it, in points of Jaccard overlap. p is the exact one-sided sign-flip test over the 10 prompts. August: 14 Aug 2026, 480 answers. October: 13 Oct 2026, 479 answers. All asked from India.
studies

Which studies has Depra AI published or planned?

One study is published, in two waves. Four more are planned, and none of them has been run at full scale yet.

  • Fieldwork or target date
    14 Aug 2026
    Engines
    ChatGPT, Gemini, Perplexity
    Answers
    480
    Data
    IndicGEO on GitHub (opens in a new tab)
  • Fieldwork or target date
    13 Oct 2026
    Engines
    ChatGPT, Gemini, Perplexity
    Answers
    479
    Data
    rerun-2026-10 on GitHub (opens in a new tab)
  • How many runs? India

    Planned
    Fieldwork or target date
    Late October 2026
    Engines
    Not set yet
    Answers
    Not set yet
    Data
    None yet
  • Google Gap

    Planned
    Fieldwork or target date
    Late November 2026
    Engines
    ChatGPT; Google results, AI Overviews and AI Mode
    Answers
    Not set yet
    Data
    None yet
  • Hinglish vs English, monthly re-run

    Planned
    Fieldwork or target date
    Monthly
    Engines
    Not set yet
    Answers
    Not set yet
    Data
    None yet
  • Search-index field test

    Planned
    Fieldwork or target date
    Not set yet
    Engines
    Search indexes such as Brave and Perplexity Search, and AI engines
    Answers
    Not set yet
    Data
    None yet
open data

Where is the research data?

In the IndicGEO repository on GitHub. It holds the raw August responses, prompts, lexicons, seeds, scripts and logs, so anyone can recompute the Study 01 numbers. The October re-run files are in its rerun-2026-10 folder (opens in a new tab).

If you want to replicate, extend or disagree with any of this, the repository takes issues and pull requests. Reach us at team@depra.ai.

ifham001 / IndicGEO

Public

What the repository holds

  • rerun-2026-10 (opens in a new tab)October re-run: 479 raw answers, the run log, lexicon sensitivity files and the comparison script
  • analyze.mjsOne command recomputes every August number
  • data/analysis/results.jsonThe August analysis output; the study page prints its SHA-256
  • CITATION.cffMachine-readable citation metadata, version 1.1.0

Code: MIT licence · Data and text: CC BY 4.0

View on GitHub (opens in a new tab)
method

How does Depra AI run its research?

Prompts and statistical tests are fixed before data collection starts, and the whole corpus is published so anyone can check it.

  • Design fixed before data

    Prompts are frozen and statistical tests chosen before collection starts. Anything we change afterwards goes into a public deviations log with the reason.

  • The engine's own noise is the baseline

    AI engines never answer the same question the same way twice. We measure rerun-to-rerun variation first, and only report an effect when it exceeds that floor.

  • Verified against ourselves

    Independent automated verification runs (AI agents working from raw data, without our analysis code) recompute the headline numbers. When they find defects in our instrument, and they have, we fix and re-run before publishing.

  • Everything published

    Raw responses, prompts, lexicons, seeds, scripts and logs. We sell AI visibility tracking, which gives us an interest in what these studies conclude. Publishing the whole corpus is how we make that interest checkable by anyone.

in progress

Which studies are in progress?

Studies Zia Wasi is preparing now. None of them has been run at full scale or published yet.

4 studies planned

Dates are targets from the research plan.

  • Planned for late October 2026

    How many runs? India

    How many repeat answers a brand needs before its AI visibility number stops moving, measured on Indian buying questions.

  • Planned for late November 2026

    Google Gap

    Why a brand that ChatGPT names can be missing from Google's organic results, AI Overviews and AI Mode, asked from India and the US, with depra.ai itself as the first case.

    Engines and indexes: ChatGPT; Google results, AI Overviews and AI Mode

  • Planned monthly

    Hinglish vs English, monthly re-run

    The same frozen prompts from Study 01, asked again each month to see how much the engines drift.

  • Planned

    Search-index field test

    A fixed set of English, Hindi and Hinglish buyer questions, sent to search indexes such as Brave and Perplexity Search and to the AI engines, to compare the pages each index returns with the pages the answers cite.

    Engines and indexes: Search indexes such as Brave and Perplexity Search, and AI engines

research notes

Where can I read shorter research notes?

Shorter write-ups from the "AI models and search indexes" research, which Zia Wasi collected from public sources.

all research posts
in use

Where are these findings used?

Our guide on how to measure AI visibility uses Study 01 to explain why English and Hinglish need separate tracking. The multilingual tracking page shows how Depra AI asks your buying questions in English, Hinglish, French and most other major languages, plus code-mixed questions, on every plan. To see how Depra AI sits next to other tools, read the AI visibility tool comparisons.

Is your brand visible in both languages?

The free AI audit asks 5 English questions on ChatGPT, Gemini, Google AI Overviews and Perplexity. Every plan tracks your brand in English, Hinglish, French and most other major languages, plus code-mixed questions. Hinglish is Depra AI's speciality.

get started

Make AI answers your next sales channel

See where ChatGPT, Gemini, Perplexity, Google AI Overviews and every other major AI engine name you today, and get a ranked list of fixes to change it.

Frequently asked questions

Who runs the research, where the data is, and how to check or challenge it.

Zia Wasi leads it. He is co-founder of Depra AI and a GEO (generative engine optimization) researcher. He is also preparing the studies listed as planned on this page.

In August 2026, asking in Hinglish changed which brands Gemini and Perplexity recommend. Beyond normal rerun noise, the brand list changed 2.2 points on ChatGPT, 7.8 on Gemini and 23.4 on Perplexity. In the October 2026 re-run (479 answers) the gap was 28.2 points on Perplexity, 2.0 on Gemini and 0.8 on ChatGPT. Read the study.

In the IndicGEO repository on GitHub (opens in a new tab). It holds the raw August responses, prompts, lexicons, seeds, scripts and logs, so anyone can recompute the Study 01 numbers. The October re-run files are in its rerun-2026-10 folder (opens in a new tab).

Depra AI sells AI visibility tracking, so it has an interest in what these studies conclude. Prompts and statistical tests are fixed before data collection starts. Any later change goes into a public deviations log with the reason. The whole corpus is published so anyone can check it.

The study "How many runs? India" is planned for late October 2026, and "Google Gap" for late November 2026. A monthly re-run of Study 01 and a search-index field test are also planned. None of them has been run at full scale or published yet.

Yes. The IndicGEO repository (opens in a new tab) takes issues and pull requests. You can also email team@depra.ai.

Be the brand AI recommends

Track your brand on ChatGPT, Gemini, Perplexity, Google AI Overviews and every other major AI engine, and get a ranked list of fixes that help AI name you.