What actually changed between August and October

On 31 August 2026, OpenAI said advertisers could buy ChatGPT ads directly through Ads Manager across India, Europe, the Middle East and North Africa. The same post reported a $1 billion annualised revenue run rate less than 200 days after launch and more than one billion weekly active ChatGPT users. By 21 October, ads were live in 89 countries, over 100,000 advertisers had run campaigns, and weekly usage was put at 1.2 billion people. Ads appear only on the Free and Go plans; Plus, Pro and Enterprise stay ad-free.

The buying interface is a recognisable performance stack: CPC and outcome-optimised bidding now account for most campaigns, with Pixel and the Conversions API underneath, plus product feeds and geo-targeting. Your feed hygiene and event setup transfers.

What it costs in India

India is the cheapest entry point on the platform. OpenAI's published minimum daily budget for Indian accounts is Rs 725, about $7.60 a day, against $25 in the United States and 15 GBP in the United Kingdom. That number matters more than any percentage rule, because it sets the floor for what a readable test costs.

Two eligibility rules decide whether you can actually spend it. Self-serve requires the advertising legal entity to be registered in a supported country, and new accounts are often limited to targeting their home market until identity verification and some home-market spend are complete. An Indian brand selling only in India is unaffected. An Indian brand exporting to the UAE or Saudi Arabia is not: neither country is on OpenAI's self-serve list.

If you are working out where this sits against the rest of your stack, our performance marketing and Google Ads pages cover how the channels fit together.

And Google did not stand still. On 19 May 2026 Google said AI Mode had passed one billion monthly users a year after launch, with queries more than doubling every quarter, and made Gemini 3.5 Flash the default model in AI Mode globally. It also made the jump from an AI Overview into a conversational follow-up live worldwide. More of the buying conversation is now mediated before anyone clicks.

The part nobody wants to put in a deck

The best public evidence on AI referral quality is a peer-reviewed study by Maximilian Kaiser and Christian Schulze in Marketing Science (45(4), 2026). It used 12 months of first-party Google Analytics data, August 2024 to July 2025, from 973 e-commerce sites with about $20 billion in combined revenue: more than 50,000 transactions from ChatGPT referrals against 164 million from traditional channels.

The finding is unfashionable. Organic traffic from large language models converted at a higher rate, and earned more revenue per session, than paid social — and lower than every other traditional channel. It was under 0.2% of all traffic in the dataset. The authors say plainly that the work is descriptive, not causal, and that last-click attribution understates a channel that mostly does discovery work. Read that as the finding: AI referral traffic is small, sits high in the funnel, and a last-click ROAS report will make it look worse than it is.

There is a second caveat. The study window closes in July 2025, before any paid placement inside ChatGPT existed. It describes organic AI referrals; it cannot tell you how an ad unit converts, because the ad unit did not exist when the data was collected. Anyone quoting it to forecast ad performance is stretching the source.

And the industry estimates disagree with it. Similarweb's 2025 ecommerce report put ChatGPT-referred visits converting at 11.4%, against 5.3% for organic search. The academics address the gap in an appendix; the difference is method, modelled panel estimates against first-party last-click GA4.

What the click itself looks like

The cleanest causal evidence on AI answers is a randomised field experiment by Saharsh Agarwal (Indian School of Business) and Ananya Sen (Carnegie Mellon), last revised in July 2026. They recruited 1,065 US participants, installed a browser extension that hid AI Overviews for a randomly assigned group, and tracked browsing for two weeks between 7 January and 10 February 2026.

When an AI Overview appeared, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%. Sponsored clicks and overall search frequency did not change. Removing the overviews produced no measurable improvement in how satisfied participants said they were with search.

Two things follow. The substitution is between the summary and the organic result, not the summary and the ad — paid placements held up. And an AI Overview appeared on roughly 41% of queries in their sample, so the audience is not a niche.

The limits: US participants, desktop Chrome only, two weeks. Do not port those percentages to India — port the direction.

Organic AI visibility and paid AI visibility are not substitutes

Our earlier playbook treated AI visibility as something you earn. That is now half the picture.

An ad buys a slot beside an answer you did not earn. It does not make the answer correct — if an assistant describes your product, price or variant wrongly, the ad sits under a wrong description and you paid for the privilege. Fix the facts first.

The reverse is also true: a well-built product page does not guarantee a mention where a competitor is paying for the slot. Two different jobs, two different budgets. Our SEO, AEO and GEO page covers the earned side in detail.

Five rules for splitting the budget

Bain's Game Changers 2026 puts India's insurgent consumer brands at $7.5 billion in FY25 revenue, 3.75 times FY20, but only 22% of those above ₹100 crore crossed ₹500 crore. Efficient growth is the constraint — hence a capped test, not a new line item.

  1. Cap the test at 3 to 5 per cent of monthly paid media, with a written stop date. At ₹40 lakh a month that is ₹1.2 to ₹2 lakh — enough for a read, small enough to survive a bad month.
  2. Buy on CPC before you buy on outcomes. You need click volume and a cost-per-click baseline in your own category before an optimiser has anything to learn from.
  3. Send every click to the page that matches the conversation. If the assistant recommended a specific variant, land on that variant, not the homepage.
  4. Instrument before you spend. Pixel and Conversions API from day one, plus a post-purchase question, because much AI influence ends in a branded Google search your attribution will credit elsewhere.
  5. Judge at 30 days on cost per first order and share of new customers, not on ROAS alone. OpenAI's examples — an unnamed advertiser at 3x return on ad spend over 28 days, a partner reporting over 80% of ad-driven traffic as new customers — are vendor claims, not benchmarks.

Worked example: a ₹2,00,000 first month

This is arithmetic, not a client result. Swap in your own inputs; the structure holds.

Say your brand spends ₹40,00,000 a month on Meta and Google. A 5 per cent test cap gives you ₹2,00,000. At an assumed ₹22 cost per click that is roughly 9,090 clicks; at an assumed 1.4 per cent on-site conversion, about 127 orders, or roughly ₹1,570 cost per order.

Compare that with contribution margin per order. If it is ₹900, the test loses money on last click — the expected month-one outcome, not a failure. The question is not "is this profitable" but "what would have to be true for this to be worth scaling".

So set the rule in advance. If cost per first order is still above ₹1,570 after 30 days and 100 orders, stop and put the money back into the channel that works. If it lands below ₹1,200 with a fair share of first-time buyers, raise the cap to 10 per cent and re-run the read. Anything between the two is noise, and you do not make budget decisions on noise.

How to measure it without lying to yourself

AI referrals arrive from the assistant's own domain, so a default GA4 setup buries them inside "Referral". Build a channel group before you spend.

Tag every link you control. Where you cannot, ask: one post-purchase question about where the customer first heard of you captures journeys that end in a branded search and get credited to organic.

Keep the manual prompt tracker running alongside the test: twenty real buying questions, weekly, recording whether you are named, whether the description is accurate and which domains get cited. It is the only signal telling you whether your paid slot sits under a true statement.

Then report AI as a share of new-customer acquisition, not a standalone ROAS line. A channel under a per cent of sessions can be important and still look trivial on a last-click report.

What we would do in month one

Week one, instrument. Pixel and Conversions API, a GA4 channel group for AI referrals, a post-purchase question on the order confirmation. No spend until those three exist, because a test you cannot read is a donation.

Week two, fix the facts. Run the twenty-prompt audit, correct anything the assistants get wrong about your product, price or variants, and align the marketplace listing with your own product page. That work is what makes a paid slot worth buying.

Week three, launch at 3 to 5 per cent with CPC bidding, one ad per product family, and landing pages that match the conversation rather than the homepage. Write the stop date and the scale threshold on the same page as the budget.

Week four, read it once. Cost per first order, share of new customers, and whether the answer your ad sits under is still accurate. Then decide: stop, hold, or raise the cap. One decision, one number.

The channel is real and it is now buyable in India. It is also small, early, and easy to overspend on if a launch announcement sets your budget instead of your own first thirty days of data.

Follow along

Follow on Instagram: @sugarycitrus.co. That is where the things that do not fit an article go — a failed test, an ad account behaving badly, a number that surprised me. Not on a schedule, only when it is worth posting.

Want this run on your brand? Send your category and monthly ad spend through the contact form. I will reply with the first three things I would fix. Free, no deck.

Sources

FAQ

Is ChatGPT Ads actually available to Indian advertisers?

Yes. OpenAI said on 31 August 2026 that advertisers could buy ChatGPT ads directly through Ads Manager across India, alongside Europe, the Middle East and North Africa. Access is self-serve.

What is a sensible first budget?

Three to five per cent of monthly paid media, held for 30 days. Month one is for a cost-per-click and cost-per-order baseline in your category, not for scale. If you cannot afford to lose that amount, do not run the test.

Does this replace any of my Google or Meta spend?

No. The evidence describes AI referral traffic as small and upper-funnel, converting above paid social but below every other traditional channel. Fund it from the experimental budget, not a working channel.

If I buy ads, does organic AI visibility stop mattering?

It matters more. The ad sits beside an answer you did not write. If that answer gets your product, price or variant wrong, you are paying to promote a mistake.

How much does ChatGPT Ads cost in India?

OpenAI's published minimum daily budget for Indian accounts is Rs 725, roughly $7.60, the lowest minimum on the platform. Treat that as a floor, not a plan. A readable test still needs 30 days and enough orders to judge cost per first order.

How long before I know whether it works?

Thirty days and roughly a hundred orders is the minimum honest read; below that you are looking at variance. Expect month one to lose money on last click — that is what the published channel data predicts for a new upper-funnel source.