What Happened When ChatGPT Started Recommending Us: An Honest AI Visibility Case Study

We are a Mumbai agency that sells SEO and GEO. So when ChatGPT started sending us visitors, we had a duty to measure it properly and every incentive to oversell it. This piece tries to do the first and avoid the second. Everything below comes from our own GA4 property, Google Search Console and the lead tracker on abovethefold.in. The numbers are small. That is part of the point.

What we saw: 201 sessions from ChatGPT in under two years

From January 2025 to 20 October 2026, GA4 recorded 201 sessions to abovethefold.in with chatgpt.com as the source. Calendar 2025 brought 99 of them, plus 17 from perplexity.ai. In 2026, monthly sessions climbed to a peak in May and then fell away sharply.

The climb from March to May was real: sessions more than quadrupled, from 7 to 31. The fall was faster, down to 2 in July. By normal traffic standards these are tiny numbers. But sessions turned out to be the wrong way to judge this channel.

Why these visitors were worth more than direct, LinkedIn or Bing traffic

Visitors from ChatGPT converted at roughly four times the rate of our direct traffic. GA4 recorded a key event rate of 4.98% for chatgpt.com sessions, against 1.24% for direct, 2% for LinkedIn and 1.82% for Bing.

For us, a key event is a call click, WhatsApp click, email click or form submission. Over the period, chatgpt.com sessions produced 20 of them. Engagement told the same story. Average engagement time was 1m 23s for ChatGPT visitors against 21s for direct, and they triggered 7.78 events per session against 4.91.

Our reading is straightforward. Someone arriving from a ChatGPT answer has already been through a shortlist: they asked a question, got a recommendation with reasons attached, and clicked to check it. It is the closest thing to a warm referral a website gets.

A caution on sample size. Twenty key events are enough to see a pattern and not enough to build a forecast on. We treat the four times figure as a strong signal, not a law.

Where ChatGPT sent people, and what they asked for

ChatGPT sent most of its visitors to a single page: our social media marketing service page. Not the homepage, and not the blog.

So social media marketing is the service ChatGPT most associates with Above The Fold, even though we offer SEO, performance marketing and web development too. And that page has to convert on its own, because AI visitors skip the homepage entirely.

The lead tracker confirms it. Since we started first-touch lead attribution on the site in late September 2026, WhatsApp enquiries with ChatGPT as the first-touch source have been recorded on 23 September, 28 September and 15 October 2026. All three landed on the social media marketing service page. If you want to build the same setup, our guide to tracking ChatGPT and AI referrals in GA4 walks through the channel group, key events and CRM fields we use.

The AI due-diligence queries we spotted in Search Console

Search Console shows that AI assistants are now running agency due-diligence searches on behalf of buyers, and some of those searches surface our pages. Our site has started receiving impressions for long, conversational queries that read like the sub-queries an assistant fires off while building an answer.

Two examples from our Search Console data:

which performance marketing agency manages reebok india’s d2c catalog ads on meta and google?

That query alone brought 20 impressions. Another:

who is the performance marketing agency behind veromoda and pepe jeans’ d2c growth on meta in india?

To be clear, we do not work for Reebok, Vero Moda or Pepe Jeans, and nothing on our site claims otherwise. The point is the shape of the query. Few people type questions like these into Google. They look like fan-out queries: a buyer asks an assistant to find the agency behind a brand they admire, and the assistant runs a string of narrow searches to assemble the answer.

AI assistants are doing reference checks. If your client categories, case studies and results are not stated plainly on crawlable pages, the assistant fills the gap with someone else’s.

Google’s AI features barely knew our service pages

Google’s AI features showed us mostly for our own name. Search Console’s generative AI report recorded 73 impressions inside AI Overviews and AI Mode in the 90 days to mid October 2026, and 68 of those were for our homepage on brand searches.

Service pages and blog posts barely appeared. Google’s AI answers know us by name but rarely pick us for category questions, such as which agency to hire for SEO in Mumbai. That gap is what our October work targets.

Microsoft Copilot cited us 676 times in six months

ChatGPT clicks were only half the picture. Bing Webmaster Tools’ AI Performance report shows that Microsoft Copilot and its partner experiences cited abovethefold.in 676 times between 21 April and 18 October 2026, usually citing one of our pages per answer.

One page did most of the work. Our guide on how to choose a digital marketing agency in India earned 303 of the 676 citations, followed by our Instagram algorithm guide and our influencer marketing pricing guide. The questions Copilot was answering when it cited us included “instagram algorithm changes india”, “SEO strategies for D2C brands organic sales growth” and “how to choose a revenue-focused digital marketing company”.

That tells us two things. First, practical buyer guides with checklists and numbers get cited far more than opinion pieces. Second, Copilot citations collapsed in October, a few weeks after our publishing pause, which is the clearest sign yet that AI visibility decays when a site stops changing. We have since refreshed the three most cited guides with new sections.

What dropped, and why we think it dropped

ChatGPT sessions fell from 31 in May 2026 to 2 in July and 4 in August. We cannot prove a single cause, but the most likely one is uncomfortable: we stopped feeding the engines.

We had been publishing regularly. Between late July and October 2026, we published no new articles at all. AI answers lean heavily on freshness, on third-party mentions, and on presence in directories and ranked lists. When our inputs stopped, our visibility dipped with them.

Being candid means adding two qualifications. First, the decline began in June, before publishing stopped, so the pause is not the whole story. AI referrals are more volatile than Google rankings: models get updated, retrieval sources shift, and the same question can return a different shortlist on different days. Some of the fall is likely noise we cannot control.

Second, analytics only sees clicks. Much of the value of an AI answer never reaches GA4. People read a recommendation, then call, send a WhatsApp message, or search our name. Branded searches such as “above the fold” and “above the fold company” are our main source of Google clicks, and we suspect some of that branded demand starts in AI answers we never see. The session count is the floor of AI influence, not the ceiling.

Exactly what we changed in October 2026

In October 2026 we went back to the inputs AI engines rely on: fresh content, structured answers, clear numbers, fast indexing and third-party presence.

• Restarted publishing. We put 11 new pages live in a single week, then moved to a twice-weekly publishing schedule.
• Added FAQ structured data across 31 posts. Question and answer pairs with FAQPage schema give engines short, quotable answers with a clear question attached.
• Added price-range tables. AI engines extract numbers. A page with clear price ranges is easier to cite than one that says “contact us for pricing”.
• Refreshed llms.txt. We updated the plain-text summary of the site that some AI crawlers read, so it reflects our current services and pages.
• Set up IndexNow pinging and Bing Webmaster Tools. Copilot is built on Bing, and IndexNow tells Bing about new and updated pages as soon as they publish instead of waiting for a crawl.
• Started on directory and list presence. We are pursuing listings on Clutch and GoodFirms and inclusion in third-party agency lists, because AI answers often cite these when building a shortlist.

What we would tell a brand trying to replicate this

Getting recommended by AI assistants is a steady supply job, not a one-time fix. Our short checklist:

1. Measure before you optimise. Group AI sources in GA4, track WhatsApp, call and form key events, and record first-touch source in your CRM.
2. Find the page AI already sends people to. Make that page convert on its own, with pricing, proof and a WhatsApp or call option above the fold.
3. State facts plainly. Who you serve, where you operate, what it costs and what results look like, written in sentences an engine can lift.
4. Publish on a schedule you can keep. A long gap in publishing costs visibility, as it appears to have done for us.
5. Get mentioned elsewhere. Directories, ranked lists, trade press and partner pages carry weight that your own site cannot.
6. Add FAQ blocks and structured data. Short, standalone answers are the easiest thing for an engine to quote.
7. Judge on enquiries, not sessions. Key event rate and first-touch leads tell you whether AI visibility is paying off.

For the full method, see our guide on how to get recommended by ChatGPT and AI Overviews, and our explainer on what a GEO agency in India actually does.

How we are measuring the next 90 days

We will judge the October changes on five measures over the next 90 days, reviewed monthly, with enquiries weighted above traffic.

If the numbers recover, we will know which inputs mattered. If they do not, that is just as useful to the brands we advise.

If you want to know whether AI assistants are recommending your brand, and what to change if they are not, talk to our team. You can also see how we approach this work on our SEO and GEO service page.