Disclosure: I'm Ifham, co-founder and CTO of Depra AI, an AI visibility tracking tool. The data below is from our own open study, and I link to the raw data so you can check it yourself.
A buyer opens ChatGPT and types "best face wash for oily skin under ₹500". The answer names two or three brands. If yours is not one of them, you never see that customer, and nothing in Google Search Console tells you it happened.
That is the problem answer engine optimization (AEO) and generative engine optimization (GEO) try to solve. This post covers what we measured across 480 AI answers, what it means for anyone shipping a website, and a short checklist you can run this week.
What are AEO and GEO?
Answer engine optimization (AEO) is the practice of structuring your content so AI assistants and answer boxes can extract a clear answer from it. Generative engine optimization (GEO) is the practice of making your brand more likely to be mentioned and cited inside AI-generated answers from ChatGPT, Gemini, Perplexity and Google AI Overviews. Classic SEO still matters, because these engines retrieve web pages, but ranking on page one is no longer the same thing as being recommended.
If you want the longer version, we wrote up the differences in AEO vs GEO vs SEO and what is AEO.
What we tested
We wanted to know whether the language of the question changes which brands an AI engine recommends. In India a lot of people type in Hinglish, which is Hindi written in Roman script and mixed with English, so this was not an academic question for us.
- 10 unbranded buying questions (5 skincare, 5 fashion), each written in English and in Hinglish
- 3 engines: ChatGPT, Gemini and Perplexity, geolocated to India
- Each prompt run 8 times, 480 responses in total
- All collected in one 57-minute window on 14 August 2026
The full write-up is on the Hinglish vs English AI shopping study page, and the data and code are on GitHub at ifham001/IndicGEO.
What we found
Gemini cited far fewer sources in Hinglish. 90.0% of English answers carried citations, against 41.3% of Hinglish answers. The average number of citations per answer dropped from 9.4 to 5.2.
Brand recommendations moved by engine. The gap between English and Hinglish recommendations was 2.2 percentage points on ChatGPT, 7.8 on Gemini and 23.4 on Perplexity (p = .0010). So the engine you are tracking changes the story completely, and an average across engines hides it.
Individual brands swung a lot. One brand went from 2 mentions to 16 on Perplexity depending on the language of the question, a swing of 17.5 percentage points.
A caveat before anyone quotes this: it is 10 prompts in two categories, in one country, in one time window. Treat it as directional evidence, not a law of nature. That is why we publish the data.
A checklist you can run this week
1. Let AI crawlers read your site
Many sites block AI bots by accident, through a blanket rule or a CDN setting. Check your robots.txt and decide deliberately. If you want to be retrievable by the major assistants, this is the starting point:
User-agent: GPTBot
Allow: /
User-agent: OAI-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Google-Extended
Allow: /
There is a full list of AI crawlers and an AI crawler checker if you want to test what your own site returns.
2. Answer the question in the first two sentences
Engines pull passages. Put the direct answer at the top of the page, then the detail. Include the category, price and key specs as plain text, not only inside images or JavaScript widgets.
3. Add structured data that matches the visible page
JSON-LD gives machines an unambiguous summary of who you are. Keep it consistent with what a human can see on the page.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "YourBrand",
"url": "https://www.example.com",
"description": "One plain sentence saying what you sell and who it is for.",
"sameAs": [
"https://www.linkedin.com/company/yourbrand",
"https://github.com/yourbrand"
]
}
</script>
4. Get mentioned where the engines already look
Your own site is only part of the picture. In our India data the most cited sources were YouTube, Reddit and Nykaa, which means third-party pages carry a lot of weight. Our guide on how to get cited by AI pulls together the research here, including an Ahrefs analysis of 75,000 brands that found brand mentions correlated with AI Overviews visibility much more strongly than backlinks did.
5. Measure per engine, per language, over time
AI answers are not stable. Sources that get cited one week are often replaced the next, and the same prompt can return a different brand list on every run. A single manual check tells you almost nothing. Run the same questions repeatedly, track each engine separately, and watch the trend rather than any one answer.
How we track this at Depra AI
Doing the measurement above by hand is slow, so we built Depra AI to automate it:
- Tracks how ChatGPT, Gemini, Google AI Overviews and Perplexity mention your brand, in English and Hinglish
- Shows which competitors appear for the same questions and which sources the engines cite
- Ranks an action plan of pages to publish and places to get mentioned
- Audits whether AI crawlers can reach your site
- Excludes questions that name your brand from the score, so the number reflects real discovery
Plans start at ₹1,999 a month with a 7-day trial, and there is an AI visibility checker if you just want a first look. See pricing for details.
FAQ
Is AEO the same as SEO?
No. SEO aims to rank pages in a list of links. AEO and GEO aim to get your brand named and cited inside an AI-written answer. They overlap, because engines retrieve web pages, but the measurement is different.
Which AI engines should I track?
At minimum ChatGPT, Gemini, Perplexity and Google AI Overviews. In our study their behaviour differed a lot, so track each one separately.
Does the language of the question matter?
In our tests it did, especially on Perplexity and for Gemini's citation rate. If your customers mix languages, test the way they actually type.
How often should I check?
More than once. Answers vary between runs and between weeks, so a repeated schedule is more useful than a single snapshot.
If you run your own tests, I would like to see the results. Drop them in the comments, and tell me where you think our method is weak.
Top comments (0)