You get cited in AI answers by being mentioned, favourably and repeatedly, on the third-party pages an engine reads while assembling its answer. Your own site matters at the margin; the bulk of the signal sits off-site, and the strongest correlate anyone has published is branded web mentions.
Ahrefs studied 75,000 brands and found branded web mentions correlate with AI citations at rho = 0.664, against 0.218 for backlinks. That single comparison should reset your budget. The work that earns citations resembles PR, reviews and community presence far more than it resembles classic link building.
Why off-site mentions do the heavy lifting
An AI engine answering "best sulphate-free shampoo under ₹800" is not ranking pages the way Google's blue links did. It pulls a handful of sources, reads the brand names inside them, and writes a shortlist. If your brand is absent from those sources, no amount of on-page work puts you in the answer.
That is why unlinked mentions count. A Reddit thread naming your brand, a review roundup, a YouTube comparison transcript, a category listicle, all feed the same pool of text, none of them needing a link to do it. The Ahrefs figure is a correlation, not proof of causation. It is still the best public signal available, and it points one way.
If your brand never appears at all, start with the diagnosis in why ChatGPT doesn't mention your brand before you spend on outreach.
Every engine reads a different slice of the web
There is no single "AI index" to optimise for. Kevin Indig and Omnia analysed 3.7 million citations across 20,000 prompts in May 2026 and found only 2.37% of cited URLs appeared in all three of ChatGPT, Perplexity and Google AI Overviews. Ninety-one percent appeared in exactly one engine.
Peec's March 2026 study of 30 million sources found Reddit the most-cited domain overall, with per-engine source preferences differing substantially beneath that. So "get on Reddit" is a reasonable default and a poor strategy on its own. The useful question is which domains get cited for your category, on your prompts, in each engine.
That is a data problem, not a guessing problem. Depra's cited-sources view lists the domains each engine actually pulled for your tracked prompts, and citation gap analysis shows the domains citing your competitors that never cite you. Those gaps are your outreach list.
The real battleground for Indian D2C
Three surface types tend to carry Indian category answers, and none of them is your homepage.
- Comparison listicles and category roundups. "Best X brands in India" posts from media sites, niche blogs and affiliate publishers. These are structurally what an engine wants: a named category, a list of brands, some reasoning.
- Review and marketplace surfaces. Amazon and Flipkart review bodies and Q&A sections, Nykaa product pages, app-store reviews. Buyers describe products there in roughly the phrasing they later type into ChatGPT.
- Community and video. India-specific subreddits in your category, long-form YouTube reviews and their transcripts, forum threads that rank well and stay up for years.
Language matters here more than most global guides admit. Indian buyers type Hinglish: "best face wash for oily skin under 500 rupees", "kaunsa protein powder theek hai". India is ChatGPT's second-largest market at roughly 100 million weekly users (Altman, Feb 2026), and Google AI Mode launched in India in June 2025 with Hindi added in September 2025. Track only polished English prompts and you are measuring a fraction of your exposure.
On-page work that helps, and the honest caveat
On-page work is cheap, so do it. The Princeton GEO paper (KDD 2024) found that adding statistics, quotes and citations to content raised visibility roughly 30 to 40% in a sandbox running GPT-3.5. Clear category framing helps for the same reason: an engine needs to know what you are before it can shortlist you for anything.
The caveat is real. C-SEO Bench (NeurIPS 2025) tested many conversational-SEO tactics and found them largely ineffective, occasionally negative. The Princeton result came from a controlled sandbox on an older model, not from live engines in 2026.
Read that as a ceiling on ambition, not a reason to skip the work. Write answer-first paragraphs under question-style headings, state your category and price band in plain text, cite sources, quote named people, keep specs out of images. Then measure whether it moved anything, and stop if it did not.
| Tactic | What the evidence says | Priority |
|---|---|---|
| Earning branded mentions off-site | rho = 0.664 with AI citations across 75K brands (Ahrefs) | Highest |
| Getting into per-engine source sets | 91% of cited URLs appear in only one engine (Indig x Omnia, May 2026) | High |
| Statistics, quotes and citations on-page | ~30-40% lift in a GPT-3.5 sandbox (Princeton, KDD 2024) | Medium, cheap |
| Generic conversational-SEO tweaks | Largely ineffective, sometimes negative (C-SEO Bench, 2025) | Low |
| Publishing llms.txt | 97% of files received zero bot requests (Ahrefs, 2026) | Skip |
What not to waste time on
llms.txt is the clearest example. Ahrefs found in 2026 that 97% of llms.txt files received zero bot requests. It takes ten minutes to publish, so publish it if you like, but do not put it on a roadmap or report it as work completed.
The second waste is treating one good answer as proof. SparkToro logged 2,961 runs from 600 volunteers in January 2026 and found the odds of the same prompt returning an identical brand list twice are under 1 in 100; identical list in identical order is roughly 1 in 1,000. A screenshot of ChatGPT naming you is a coin flip, not a result. The distinction between optimising for answers and optimising for rankings is covered in AEO vs GEO vs SEO.
Citations churn, so treat this as maintenance
Even when you win a citation, it does not stay won. SISTRIX found ChatGPT replaces around 74% of its cited sources week over week. Kevin Indig's June 2026 analysis of 815,000 prompt-page pairs found only 2.2% of citations persisted across three ChatGPT runs, with within-model variance of 10 to 34%.
So a one-off PR push produces a spike you cannot see and cannot hold. What compounds is a steady flow of new mentions across the source types each engine favours, measured over months rather than checked after a campaign.
The prioritised action list
- Measure your baseline properly. Fixed prompt set, English and Hinglish, all four engines, enough runs to produce a rate rather than an anecdote. Record the sample size alongside every score.
- Pull your cited-sources list per engine. Find which domains each engine already trusts for your category prompts.
- Run a citation gap analysis against two competitors. The domains citing them and not you are your target list, ranked by how often the engines actually pull them.
- Get into category roundups and comparison listicles. Pitch publishers already cited for your prompts. One inclusion in a roundup the engines already pull is worth more than ten posts nothing cites.
- Fix your review surfaces. Volume and recency of reviews on Amazon, Flipkart, Nykaa and your own product pages, plus resolving the complaints that keep resurfacing.
- Show up in community and video. Answer real questions in your category honestly, where your buyers already read. Astroturfing is both detectable and reputationally expensive.
- Do the on-page hygiene once. Category framing, statistics, quotes, citations, answer-first structure. Then leave it alone and watch the numbers.
- Re-measure monthly and act only on moves outside the confidence band. Daily noise will otherwise send you chasing ghosts.
How to tell whether it is working
Because engine output is this noisy, a visibility number without a sample size is decoration. Depra ships every visibility score with its sample size and a 95% Wilson confidence interval, and only alerts on changes that fall outside that band. Prompts run daily across ChatGPT, Google AI Overviews, Gemini and Perplexity, Perplexity weekly on paid plans, with visibility split by English and Hinglish, plus competitor benchmarking on identical prompts, sentiment, average position and CSV export.
The free plan tracks 5 prompts weekly across ChatGPT, Gemini and AI Overviews, enough to see whether your baseline is zero or something. Paid plans start at ₹1,999/mo excluding GST, with INR billing and GST invoices; the full breakdown is on pricing, signup is self-serve and the first scan returns results in minutes, and a comparison of the alternatives is in the best AI visibility tools for D2C.
Whatever you measure with, hold the prompt set fixed. Change prompts mid-quarter and every before-and-after comparison dissolves: with the engines moving this much on their own, you will have no way to separate your change from theirs.