We Tested 4 AI Engines on 15 Local Business Searches

Every local business in India is now being told to "optimise for AI search". Almost nobody can tell you what the assistant actually reads before it names someone.

So we ran the test. We took 15 real local buyer questions - the kind a person types when they need a root canal in Koramangala, an emergency plumber in Andheri East at 9pm, or a CBSE school near Noida Sector 62 - and put each one to ChatGPT, Gemini, Perplexity and Claude with live web search enabled. Sixty answers. Every citation logged. Twelve questions about Indian cities, and three identical questions rewritten for Austin, Manchester and Dubai as a control.

The headline is not which engine is smartest. It is this: in India, 37% of the sources these assistants read for a local query were third-party directories. On the identical questions asked about Austin, Manchester and Dubai, that figure was 6%. The buyer asking about your business is being answered from someone else's page, and that someone else is usually Justdial.

Here is the full dataset, the ranking, and what we would change first if we owned a local business today.

The short version

• All four engines will name businesses - except when Claude opts out. ChatGPT, Gemini and Perplexity ran a live search and named specific businesses on 15 of 15 questions. Claude declined on 5 of 15 and referred the buyer to directories instead.
• India's local AI layer is a directory layer. 37% of unique cited domains in Indian answers were third-party platforms, against 6% internationally. On the emergency plumber question the gap was 69% versus 6%.
• Justdial is the most-read page in Indian local AI search. It appeared in 25 of 48 Indian answers. It appeared in zero of the 12 international ones.
• The engines rarely agree. On 11 of 15 questions there was no single business that all four assistants named.
• Perplexity gives you the most room; ChatGPT the least. 13.1 unique domains per answer versus 4.9.
• Being cited is not being recommended. Perplexity pulled our own site into its sources for a Kolkata agency query and then named four other agencies.

How we ran it

Fifteen prompts, written the way people actually talk to an assistant rather than the way they type into Google. Twelve covered Indian cities and verticals with real commercial weight: a dentist in Koramangala, an emergency plumber in Andheri East, a gym in Gurgaon Sector 29, a chartered accountant in Kolkata, interior designers in Pune, a work-meeting cafe near T. Nagar, a car service in Gachibowli, packers and movers in Ahmedabad, an IVF clinic in Jaipur, CBSE schools near Noida Sector 62, a wedding photographer in Goa, and an SEO agency in Kolkata.

Three more were exact mirrors of the first three, relocated: the same root-canal question about Austin, the same emergency-plumber question about Manchester, the same personal-training gym question about Dubai Marina. Those three exist so that any India-specific pattern can be checked against a non-India baseline rather than assumed.

Each prompt went to four engines with web search on, through the DataForSEO AI Optimization API on 1 September 2026: ChatGPT (gpt-5.4), Gemini (gemini-3.5-flash), Perplexity (sonar-pro) and Claude (claude-sonnet-5). For every answer we recorded the full text, every citation the engine attached, the domain behind each citation, and - where the engine exposes it - the sub-queries it ran to build the answer.

That produced 60 answers, 1,050 citations and 302 unique cited domains. Every domain was then hand-classified into one of five buckets: third-party platform (a directory, marketplace or booking app), the business's own website, user-generated content (Reddit, Quora, Facebook), editorial publisher, or official body. The classification map is a checked-in file, not a judgement call made per row, so the same domain is always counted the same way.

One deliberate limitation up front: this is a single snapshot, one run per prompt per engine. It measures what these four assistants did on one day, not how stable they are week to week. We have measured that instability separately over 90 days, and it is considerable.

Finding 1: Three engines answer. One steps back.

ChatGPT, Gemini and Perplexity searched the web and named specific businesses on every one of the 15 questions. Claude searched on 10 and declined on 5.

The declines were not random. Claude opted out on the emergency plumber in Mumbai, the chartered accountant in Kolkata, the SEO agency in Kolkata, the dentist in Austin and the plumber in Manchester. Its response was consistent and, in fairness, defensible:

I don't have access to a live directory of local businesses or real-time booking information, so I can't recommend a specific plumber in Andheri East with certainty about their current availability today.

It then listed five places to look: Google Maps, Urban Company, Justdial, Housejoy or Sulekha, and local Facebook or WhatsApp community groups.

For a local business owner this is worth reading carefully, because it changes what you are optimising for. On three engines you are competing to be named in the answer. On Claude, a third of the time, you are competing to be the top listing on the directory the assistant sends the buyer to. Those are different jobs, and the second one is still classic local SEO and Google Business Profile work rather than anything new.

Finding 2: In India, the assistant reads a directory. Abroad, it reads the business.

This is the finding that reframes the whole exercise.

Across 45 Indian answers that ran a search, the engines cited 363 unique domain slots. 135 of them - 37% - were third-party platforms: Justdial, Practo, Sulekha, Urban Company, NoBroker, Porter, Magicpin, Houzz, GoMechanic, Ezyschooling. Only 54% were the businesses' own websites.

Across the 10 international answers that ran a search, the engines cited 101 unique domain slots. Six of them - 6% - were platforms. 83% were the businesses' own websites.

The mirrored prompts make the comparison airtight, because the question is word-for-word identical apart from the city.

The uncomfortable reading of that chart is that this is not something the engines are doing to Indian businesses. It is a straightforward consequence of what is available to read.

Look at what Austin produced. Asked for a root-canal dentist, ChatGPT cited eaofaustin.com, austincenterendo.com, endoaustin.com, the American Association of Endodontists and the Texas State Board of Dental Examiners. Five sources, four of them practices' own sites, and the answer was built out of what those practices say about themselves. Perplexity's version of the same question cited fourteen domains, eleven of them individual practices.

Now look at Koramangala. Same question, same engines. Perplexity's sources were Justdial category pages for "Dentists For Rct Root Canal in Koramangala", Justdial pages for "Dental Surgeons for Specialized Root Canal", and a handful of clinic sites. The clinics exist. Their websites exist. They just do not answer the question as completely as the directory page does.

That is the whole gap in one sentence: the Austin endodontist publishes its own operational detail, and the Koramangala clinic outsources it to Justdial. One of those businesses gets quoted by name from its own page. The other gets included in a list someone else owns.

Finding 3: Justdial is the most-read page in Indian local AI search

Justdial appeared in 25 of 48 Indian answers - more than half - and in zero of the 12 international ones. It was cited more than twice as often as the next platform.

The per-engine split is instructive:

Claude's zero is not a preference. It is a side effect of searching less: it ran a search on only 9 of 12 Indian questions and cited an average of 3.5 domains when it did.

Behind Justdial the pattern is vertical-specific rather than general. NoBroker and Porter and AssureShift dominated the packers-and-movers answers. Practo carried the medical ones. Ezyschooling carried schools. Houzz carried interiors. GoMechanic and AckoDrive carried car servicing. Reddit turned up in 13 of 48 Indian answers and 3 of 12 international ones - lower than its reputation in AI-citation discussions would suggest for local queries specifically, but not negligible.

The actionable version of this finding is narrow and boring and worth doing this week. There are two or three platforms that decide your category. Not fifteen. Find out which ones the assistants actually cite for your service in your city, and treat those listings as owned media rather than as a directory chore you completed in 2019.

Finding 4: The engines mostly do not agree on who is best

Across 15 questions, there were only 4 where all four engines named the same business.

Three of those four were chains: Anytime Fitness for the Gurgaon gym question, Indira IVF for the Jaipur fertility question, Fitness First for the Dubai Marina gym question. The fourth was the Noida schools question, where two Sector 62 schools - Indus Valley Public School and ASPAM Scottish School - appeared in all four answers, which is what happens when a micro-locality contains a small, heavily documented set of institutions.

On the other 11 questions, no single business was named by all four assistants.

The Kolkata agency question is a clean example. ChatGPT named one set of agencies sourced almost entirely from GoodFirms and Clutch. Gemini named a different set. Perplexity named a third set. Claude named nobody. Exactly one agency appeared in three of the four answers.

Two things follow from this. First, there is no single "AI ranking" to win, which makes the whole framing of AI search as a leaderboard misleading. Second - more usefully - low agreement means each engine is a separate and individually winnable surface. You do not need to be the consensus best business in your city. You need to be one of the three to thirteen sources a given engine happened to pull for a given question, which is a much lower bar than it sounds and a much more specific target than "rank higher".

Finding 5: How much room each engine gives you

The single most useful number for a local business is not how many citations an engine attaches. It is how many businesses' own websites it will pull into one answer, because that is the number of seats available to you.

Perplexity reads widely and cites everything, which makes it the most forgiving surface for a business that has never done this work. ChatGPT reads narrowly - under five domains per answer - which makes it the hardest to enter and the most valuable slot to hold.

ChatGPT compensates with the most legible retrieval behaviour of the four. It ran an average of 6.9 sub-queries per answer, and those sub-queries are visible in the response data. For the Koramangala dentist question they included "best dentist root canal Koramangala Bangalore endodontist", "Koramangala Bangalore root canal dentist reviews" and "Practo Koramangala dentist root canal Bangalore". Gemini averaged 3.8 sub-queries, Claude 1.1, and Perplexity exposed none at all.

Read that ChatGPT list again, because it is the most actionable single artefact in this study. The engine is not searching your head keyword. It is searching service plus procedure plus micro-locality, and it is explicitly searching for the directory page alongside the general query. If your website has one page called "Dental Services" and no page that names the procedure and the neighbourhood, you are not in that retrieval set - and no amount of technical SEO work fixes a page that does not exist.

Finding 6: Being cited is not being recommended

We put our own category in the test set: "which digital marketing agency in Kolkata should I hire for SEO".

Perplexity pulled nicodigital.com into its source list for that question. It then named four other agencies in the answer and did not mention us once.

That is worth sitting with, because we run this work for a living and we still lost that answer. Retrieval and recommendation are different states. Retrieval means the engine judged your page relevant enough to open. Recommendation means the page gave it something concrete enough to justify putting your name in a sentence a stranger will act on.

The gap between the two is almost always specificity. Every business that got recommended with a reason attached in our 60 answers had the reason published on a page it controlled: the clinic that says it does microscopic root canal retreatment, the gym that publishes its 24-hour access policy, the school that lists its board and its Sector 62 address. The engines were not inferring those facts. They were quoting them.

If you are tracking AI citations, track the two states separately. A rising citation count with a flat mention count means your pages are readable and not quotable, and the fix is editorial, not technical.

The ranking

Ordered by how much opportunity each engine offers a local business that decides to do this work properly.

Two caveats on that ranking. It measures opportunity, not audience size - if your buyers use ChatGPT and nothing else, ChatGPT is your priority regardless of where it sits in this table. And it reflects one snapshot on one day; engine behaviour moves, and Claude in particular is one product decision away from searching on every local query.

What we would fix first

Five moves, in the order we would actually do them.

1. Publish the operational facts on your own site. Hours including Sunday. The exact street address with the locality name spelled the way locals say it. Phone number in text, not in an image. The procedure or service named specifically - "microscopic root canal retreatment", not "advanced dental care". Who performs it and what their qualification is. Price band if you can bear it. Every recommendation-with-a-reason in our dataset came from a page that had these facts in crawlable text.

2. Build one page per service per locality. ChatGPT's sub-queries were service plus procedure plus neighbourhood. That is the page you need to have. Not a city page, not a services list - a page whose H1 contains the thing being searched for. This is unglamorous local SEO that happens to be exactly what the AI retrieval layer rewards, whether you run a dental practice, a plumbing business, a gym or a small local service business.

3. Own the two or three platforms that decide your category. For most Indian local businesses that means Justdial and your vertical's specialist - Practo, Ezyschooling, Houzz, NoBroker, GoMechanic, GoodFirms. Complete listing, correct category, accurate hours, real photos, review volume. These are the pages the assistant reads before it reads yours, and in India it will read them more often than not. Google Business Profile remains the foundation under all of it.

4. Earn corroboration outside your own domain. Reviews with volume and recency, a genuine Reddit or Quora thread in your category, a mention in a local publication. This is where entity and knowledge-graph work starts paying - the engine needs to verify you from somewhere that is not you.

5. Check the crawlers can get in. GPTBot, PerplexityBot, ClaudeBot, Google-Extended. If your robots.txt blocks them, none of the four steps above matter. We have written separately on whether blocking AI crawlers is ever the right call; for a local business the answer is essentially never. While you are in there, confirm your schema markup carries LocalBusiness fields with a real address and real opening hours.

Everything in that list is cheap relative to paid acquisition, and none of it requires believing anything about where AI search goes next. It is the same work whether the buyer arrives from an assistant, an AI Overview, or a plain blue link.

What this study cannot tell you

Fifteen questions is a probe, not a census. It is enough to establish that the India-versus-international directory gap is large and consistent across three verticals, and not enough to put a confidence interval on it.

We ran each prompt once per engine. Our 90-day citation study found that repeated identical queries disagree with themselves a substantial share of the time, so treat every individual name in this dataset as one draw rather than a ranking. The aggregate patterns - directory share, source counts, search-abstention - are far more stable than any single answer.

We tested one model per engine. A different model tier within the same product may search more or less aggressively. And we tested through an API rather than the consumer apps, which have their own location handling and personalisation.

Finally, we are an agency writing about a market we sell into. The correction for that is to publish the method and the numbers, which is why the prompt set, the runner script and the domain-classification map are all in the repository behind this site rather than described vaguely in prose.

Where to go next

If you want the conceptual map before the tactics, start with SEO vs AEO vs GEO and what answer engine optimisation actually is. If you want engine-specific tactics, we have detailed guides on how to rank on ChatGPT and how to rank on Perplexity. If you want the broader adoption numbers this sits inside, the AI search statistics deck is updated quarterly, and our audit of 50 D2C brands covers the same question for consumer brands rather than local businesses.

For the strategic case - why brands ranking on Google are still invisible in assistants - the AI search gap is the piece to read, and what our own GA4 showed after a year of AI search traffic covers the measurement side.

And if you would rather have someone run this test on your category and hand you the fix list, that is what our AI SEO and SEO audit engagements do. Kolkata and Mumbai businesses can start with our digital marketing agency in Kolkata page for local context.

The window here is genuinely open. In 11 of 15 questions the four assistants could not agree on a single business - which means in most Indian local categories, nobody has won this yet.