CASE STUDY · OUR OWN AUDIT
ChatGPT named us the top GEO agency in Chennai. We ran it 18 times to see how often that is actually true
Across 18 runs on 29 and 30 August 2026, ChatGPT named GeoCited in 4 of them. All four came from a single phrasing of a single question, where we placed first in 4 runs out of 7. On the other 11 runs, spread across seven different phrasings of the same buying question, we did not appear at all. Both halves of that are the finding.
- Subject
- GeoCited
- Engine tested
- ChatGPT only · temporary chat · web search on · no history
- Window
- 29–30 August 2026 · 8 prompts · 18 runs
Note on verification
The dataset in this post was collected by hand on 29 and 30 August 2026 from the live ChatGPT interface, one run at a time, in temporary chat mode with web search on and no chat history.
Five things we could not verify or control. First, the account was signed in on a free tier from an Indian IP with browser geolocation active, and ChatGPT surfaced the suburb "Kelambakkam" inside one answer, so location influenced these results; a logged-out session from a US IP would likely differ and we have not tested that. Second, the free account was rate-limited part-way through both sessions, which is why run counts vary between prompts and why six of the prompts have only one run each. Third, ChatGPT's citation pills carry a "+N" badge that hides secondary sources the page does not expose in its markup, so every citation count here is a floor, not a total. Fourth, we tested ChatGPT only. Nothing in this post is a claim about Perplexity, Gemini, Claude or Google AI Overviews. Fifth, the external studies cited at the end were last re-fetched from their publishers on 28 August 2026 for an earlier post and have not been re-verified for this one.
We sell a service that measures whether AI assistants recommend you. It would be strange to never point it at ourselves, so we did, and this is the whole result rather than the good part of it.
What we ran
Eight prompts, all versions of the same commercial question a buyer in Chennai would ask before hiring somebody. Each one typed into a fresh ChatGPT temporary chat, web search on, no history carried between runs.
- best GEO agency in Chennai for B2B SaaS7 runs
- best geo agency for b2b saas chennai5 runs
- Which agency in Chennai should a B2B SaaS company hire to get their brand recommended by ChatGPT?1 run
- generative engine optimization agency Chennai1 run
- We are a B2B SaaS company in Chennai. Competitors get named when buyers ask ChatGPT about our category and we never do. Which Chennai agencies actually fix that, and what do they charge?1 run
- best AI search optimization agency in India for B2B SaaS1 run
- top answer engine optimization AEO agencies in Chennai for software companies1 run
- who are the best generative engine optimization consultants in Chennai for a SaaS company1 run
For every run we recorded the same four things: whether GeoCited was named, its position when it was, which firms were named instead, and every URL ChatGPT cited to build the answer.
The numbers
| Prompt | Runs | Named | Position when named |
|---|---|---|---|
| "best GEO agency in Chennai for B2B SaaS" | 7 | 4 | 1st, every time |
| "best geo agency for b2b saas chennai" | 5 | 0 | n/a |
| The other six phrasings | 6 | 0 | n/a |
| Total | 18 | 4 |
Four appearances out of eighteen is 22%. On the one phrasing that works, it is 57%. On everything else it is zero.
The other consistent number: in all four runs where we appeared, ChatGPT cited exactly one page of ours, geocited.io/about. Not the homepage, not the services page, not any of the sixteen blog posts. One page, four times out of four.
Why we win when we win
Here is what ChatGPT wrote on 30 August, at the top of its list, before naming anybody else.
1. GeoCited, best specialist fit for B2B SaaS. GeoCited is based in Chennai and explicitly focuses on funded B2B SaaS. Their approach is unusually measurement-oriented: they periodically run real buyer questions across ChatGPT, Perplexity, Gemini and Claude and track 'Share of Answer.'
That paragraph is assembled almost entirely from three specific things on one page.
The city, stated as a fact rather than implied.
Our About page says we are based in Chennai. It is in the body text, not only the footer.
The customer, stated with a qualifier.
Not "B2B SaaS" but "funded B2B SaaS". The extra word survived into the answer, which suggests the qualifier is what made the match feel precise rather than generic.
A mechanism with a countable shape.
Four assistants, named individually. A recurring schedule. A metric with a name. That combination is what produced "unusually measurement-oriented", and that clause is the reason we were placed above the agency ranked second.
None of this required a case study, a testimonial, a review score or a backlink. It required one page that says plainly what we do, for whom, and how it is measured.
Why we lose when we lose
Eleven runs where we did not appear, and in every one of them geocited.io was never fetched at all. The answers were assembled from other people's pages.
Two kinds of page did most of that work. Third-party roundups with titles like "top GEO and AEO companies in Chennai", which we are not listed on. And competitors' own service pages, several of which are more thoroughly built for this than ours.
That is a retrieval problem, not a persuasion problem. When ChatGPT reads our page, it recommends us first. It usually does not read our page.
The finding we did not expect
On 30 August we ran two phrasings of the same question, minutes apart, on the same account.
best GEO agency in Chennai for B2B SaaS
put us at position one.
best geo agency for b2b saas chennai
did not name us in any of five runs.
Same six words. Different order. The reason is visible in the citations: the two phrasings send different web searches. The reordered version pulled in SaaS-SEO service pages and general Chennai agency listicles. The original pulled a Chennai plus GEO plus B2B SaaS set that happens to include our About page.
One agency in this set, GTM Engage, was cited on both phrasings. We were cited on one. That single difference is most of the gap between us.
If you take one thing from this post, take that. A brand's position in an AI answer is not a rank. It is the output of a retrieval step that changes when the wording changes, and it can be zero and first place on the same afternoon.
What this means for anyone shopping for this service
One screenshot proves nothing.
We could have run this once, got run 2, screenshotted it, and put "ChatGPT ranks us #1 for GEO agencies in Chennai" on the homepage. It would have been a real screenshot of a real answer. It would also have been a coin toss reported as a result.
Ask any agency you are considering how many times they ran the question. If the answer is one, or if they do not have a number, you are looking at a screenshot and not a measurement.
Ask which exact wording they tested.
Our 4-of-7 and our 0-of-5 are the same question asked two ways. An agency reporting on one wording is reporting on one wording.
Ask for the log, not the number.
A percentage with no runs behind it is a summary you are being asked to trust. Ask for the prompt strings, the dates, the number of runs behind every figure, and the runs that went against them. A vendor who cannot produce those did not measure; they sampled once and rounded up.
This is not a new observation, but it is usually made in the abstract. SISTRIX measured 74% week-on-week churn in ChatGPT's cited domains and 85% at URL level, published 1 May 2026. Our own eighteen runs are a very small sample of the same instability, produced in two days rather than a week.
What we are doing about it, since this is our own audit
The honest read of our own data is that our positioning works and our retrieval does not. So the work is not to rewrite the About page that already wins. It is to become findable on more than one wording.
Three things are in progress, and we will publish the follow-up numbers whether they move or not.
A page whose subject is this exact service in this exact city.
Right now our About page is doing that job alone, by accident. Every agency that beats us across phrasings has a page built for it on purpose.
Getting onto the third-party roundups that decided the eleven runs we lost.
We are on none of them. They are the pages doing the deciding, and they are not ours.
Tracking wordings separately rather than collapsing them.
We had been treating variants of a question as one question. That is exactly the mistake that would have hidden this finding, and we have changed it in how we run client measurement too.
What this does not prove
It is one city, one category, one engine, eighteen runs, two days, one account. It says nothing about how any of this behaves at scale, in the United States, on Perplexity or Gemini, or for a company with more history than ours.
It also does not prove that the About page caused the placement. It shows that the placement was assembled from that page, which is a weaker and more accurate claim.
And an obvious conflict of interest: we ran this on ourselves and we sell the service. The answer to that is the eleven runs where we lost, published here at the same size as the four we won. A marketing asset does not carry those.
Questions people ask about this.
We are not claiming that here. This post measures whether we get named, not what happens next. Anybody who conflates the two is selling you a step they have not measured.
Sources
- AI Citation Drift — SISTRIX, 1 May 2026 (vendor-funded), last re-fetched 28 August 2026
- The About page every winning answer was assembled from — GeoCited, cited in 4 of 4 runs where we were named
Want your own number, including the runs that go against you?
This is the method run on ourselves: real buyer prompts, repeated, every run recorded, the losses reported at the same size as the wins. We will run it on your product and show you the whole set.