Contact Us

Work VBS Research · AI visibility

Ranking on Google isn’t being named by ChatGPT

Why SEO isn’t AEO, what Sri Lankan businesses keep getting wrong, and why the answer you see on your own phone is the least reliable one of all.

VBS Research27 October 202610 min read

A family in Colombo planning a weekend in Galle used to type “boutique hotels Galle fort” into Google and scroll. Now a growing share of people open ChatGPT, Gemini or Meta AI inside WhatsApp and ask: “Where should we stay in Galle with two kids, under Rs 40,000 a night?” They get one answer. Three or four hotels, a reason for each, and maybe a link.

That changes what “being found” means. When Google shows an AI summary, people click a regular result link far less often: in 8% of visits, against 15% when no summary appears1. The answer itself has become the shop window.

Most Sri Lankan businesses we talk to are still optimising for the old window. Some are paying for “AI SEO” packages that change nothing. And almost all of them are checking their progress in a way that can’t work.

The difference

SEO gets you into the pool.
AEO is about being picked.

QuestionSEOAEO
The question you askWhere do I rank for this keyword?Am I in the answer, and is it right?
What the customer seesTen links, ads and a mapOne answer naming a few businesses
Who decidesOne search engine’s rankingSeveral assistants, each pulling from many searches at once2
What helpsRelevance, links, a healthy siteFacts an assistant can find and quote, told the same way across trusted sources
How you measureRank and clicks for a keywordHow often you’re named, recommended, cited and described correctly
What a result looks likeA position that holds for weeksA probability that shifts with every model update

This is not a story where SEO is dead. When researchers tested popular “conversational SEO” tricks across many topics and competitors, most of them didn’t work, and some made things worse; classic SEO strategies did better3. Google says the same from its side: its AI features are built on the same search foundations, and you don’t need special files, markup or rewrites to appear in them4.

But a page one ranking is not an answer. An assistant runs several searches behind one question, reads a mix of sources, and decides who to name and how to describe them2. You can rank first for “POS system Sri Lanka” and still never be recommended, or be recommended with last year’s prices. That gap is what AEO works on.

What goes wrong

Six mistakes we see
in Sri Lanka every week

These businesses are real, but we’ve changed their names to protect their privacy.

  1. 01

    A boutique hotel in Galle

    Checking ChatGPT on your own phone

    The owner of “Ceylon Crest” asks ChatGPT for the best boutique hotels in Galle. Ceylon Crest comes up first. Job done. Except ChatGPT remembers details from her past chats5, including the dozens where she drafted Ceylon Crest’s menus and replies to guests, and it uses her location to shape local results6. A family in Dubai asking the same question gets a different list. Ask twice, and even the same person gets a different answer: identical requests to ChatGPT regularly come back different7.

    The fix: stop trusting one screenshot. Measure from a clean start, on several assistants, many times, the way your customers actually ask.

  2. 02

    A bakery in Kandy

    Treating a Facebook page as a website

    “Hela Bakes” has 40,000 followers, a price list posted as a photo, and opening hours in a pinned comment. When someone asks an assistant for a birthday cake in Kandy that can deliver by Saturday, there is almost nothing for it to read or quote. The bakery with a plain, two-page website listing cakes, prices, delivery areas and hours gets named instead.

    The fix: a simple, fast site in plain text with the facts customers ask for. Google’s advice for its AI features is the same as for search: helpful, people-first content it can crawl and index8,2.

  3. 03

    An IT company in Colombo

    Sharing a name with someone else

    There are “Apex Solutions” in Colombo, Chennai and Dubai. Ask an assistant who Apex Solutions are, and it may blend all three: the Chennai office’s reviews, the Dubai firm’s services, a phone number that rings in India. Sri Lanka is full of shared and generic names, and we see assistants confuse companies far more often than owners expect.

    The fix: describe the business the same way everywhere: website, Google Business Profile, directories, LinkedIn, press. Organisation markup that ties those profiles together helps search engines connect them9, but it supports the work rather than replacing it: in one large study, adding schema barely moved AI citations10.

  4. 04

    A tuition institute in Nugegoda

    Buying the myths

    The institute paid for an “AI SEO package”: an llms.txt file, FAQ schema on every page, and forty blog posts stuffed with statistics and quotes. Six months later, nothing changed. That’s not surprising. Google says its search ignores llms.txt files and needs no special markup for AI features4, the file itself is only a proposal11, and Google retired its FAQ rich results in May 202612. The original research on “generative engine optimisation” did find gains, but in a controlled benchmark, not a promise for every engine13; later tests found most generic rewrites ineffective3.

    The fix: ask any provider for the baseline, the questions, the assistants, the number of runs and the margin of error. If the answer is a single “AI score” and a before-and-after screenshot, walk away.

  5. 05

    A solar installer in Kurunegala

    Hiding the facts customers ask for

    “SunPeak Solar” lists “call for price” everywhere. Customers ask assistants what a 5 kW system costs in Sri Lanka and which installers are good value. The assistants answer with the installers who publish price ranges, warranties and what affects the price, and politely note that SunPeak’s pricing isn’t public. In our tests, assistants regularly flag missing prices as a weakness in their own words.

    The fix: publish ranges, what drives them, and what’s included. Microsoft’s guidance for its AI answers asks for exactly this: clear, specific, up-to-date information backed by evidence14.

  6. 06

    A law firm in Colombo 03

    Blocking the wrong robots

    The firm’s hosting company turned on an aggressive firewall that challenges every unknown bot. It also blocks the crawlers AI assistants use to search the web. OpenAI, Anthropic and Perplexity each run separate bots for search and for training, and blocking the search bot keeps you out of their answers15,16,17. Many owners block “AI” wholesale without knowing which is which.

    The fix: decide search access and training access separately, then check the firewall actually lets verified search crawlers through.

Why self-checks fail

Same question.
Very different answer.

Ceylon Crest’s owner asks “best boutique hotels in Galle” fifty times on each assistant from her own accounts. Then we ask it fifty times each the way a stranger would: fresh session, no history, a Sri Lankan location. Each dot is one answer; lit dots name Ceylon Crest.

ChatGPT16%
Gemini10%
Claude4%
Meta AI18%

Asked like a customer, Ceylon Crest is named in 12% of answers. Competitors with clearer facts and more independent mentions take the rest.

VBS Research

How we measure
what customers actually see

  1. 01

    Ask the way customers ask

    Not “Is Ceylon Crest good?” but the questions people really type: by need, by budget, by problem, in English, Sinhala, Tamil and the Singlish in between. Assistants handle lower-resource languages less reliably18, so answers in Sinhala can differ from answers in English.

  2. 02

    Ask where customers ask

    Across the assistants the public uses, not one tool. Every question starts clean: no memory, no history, no account that knows who you are, set to the market your customers are in.

  3. 03

    Ask again, and again

    Answers vary from one run to the next7, so one answer tells you almost nothing. We repeat each question until the numbers settle, and report a rate with an honest margin of error, not a screenshot.

  4. 04

    Named is not recommended

    We track separately whether you’re mentioned, recommended, cited as a source, and described correctly. A citation is not an endorsement, and it doesn’t prove the answer is right: in one audit, only about half of generated sentences were fully supported by their citations19.

What stays in the vault

The parts that make it work are ours

How we design the question panel, how answers are scored and checked, how we model the uncertainty, and how the findings turn into a ranked list of fixes for your business. That’s built on our own research library and benchmarks across the major assistants, and it’s why our clients get a plan instead of a checklist.

Questions

Before you
ask us

Is AEO replacing SEO?

No. Search-connected assistants still rely on being able to find and read your pages, so good SEO stays the foundation4,3. AEO adds a different goal and a different way of measuring it.

Can’t I just ask ChatGPT myself?

You can, but you’ll see your own personalised answer, shaped by your history and location5,6. It’s the one answer your customers will never get.

Do I need llms.txt or special schema?

Not for Google’s AI features4. Clear facts, accessible pages and consistent descriptions across trusted sources matter more than any single file.

How long does it take?

There’s an audit first, then fixes, then monthly measurement. Mentions usually build over a few months, not overnight, and anyone promising a guaranteed jump is guessing.

References

The research behind it

Every claim about how AI search works builds on published research or official documentation from the companies that build these systems. Numbers in the text point here.

  1. Chapekis, A., & Lieb, A. (2025). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center. www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
  2. Google Search Central. AI features and your website. Google for Developers documentation. developers.google.com/search/docs/appearance/ai-features
  3. Puerto, H., Gubri, M., Green, T., Oh, S. J., & Yun, S. (2025). C-SEO Bench: Does conversational SEO work? Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track. papers.neurips.cc/paper_files/paper/2025/file/27aa3aeff0f8460a7b43d30fa6c5c032-Paper-Datasets_and_Benchmarks_Track.pdf
  4. Google Search Central. (2026). Optimizing your website for generative AI features on Google Search. Google for Developers documentation. developers.google.com/search/docs/fundamentals/ai-optimization-guide
  5. OpenAI. Memory in ChatGPT. OpenAI Help Center. help.openai.com/en/articles/8590148-memory-in-chatgpt
  6. OpenAI. Searching the web with ChatGPT. OpenAI Help Center. help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt
  7. Ouyang, S., Zhang, J. M., Harman, M., & Wang, M. (2025). An empirical study of the non-determinism of ChatGPT in code generation. ACM Transactions on Software Engineering and Methodology, 34(2). doi.org/10.1145/3697010
  8. Google Search Central. Creating helpful, reliable, people-first content. Google for Developers documentation. developers.google.com/search/docs/fundamentals/creating-helpful-content
  9. Google Search Central. General structured data guidelines. Google for Developers documentation. developers.google.com/search/docs/appearance/structured-data/sd-policies
  10. Ahrefs. (2026). We tracked 1,885 pages adding schema. AI citations barely moved. Ahrefs Blog. ahrefs.com/blog/schema-ai-citations/
  11. Howard, J. (2024). The /llms.txt file (proposal). llmstxt.org. llmstxt.org/
  12. Google Search Central. Latest Google Search documentation updates. Google for Developers. developers.google.com/search/updates
  13. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 5–16. doi.org/10.1145/3637528.3671900
  14. Microsoft Advertising. (2025). Optimizing your content for inclusion in AI search answers. about.ads.microsoft.com/en/blog/post/october-2025/optimizing-your-content-for-inclusion-in-ai-search-answers
  15. OpenAI. Overview of OpenAI crawlers. OpenAI developer documentation. developers.openai.com/api/docs/bots
  16. Anthropic. (2026). Does Anthropic crawl data from the web, and how can site owners block the crawler? Claude Help Center. support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler
  17. Perplexity. Perplexity crawlers. Perplexity documentation. docs.perplexity.ai/docs/resources/perplexity-crawlers
  18. Ahuja, K., Diddee, H., Hada, R., Ochieng, M., et al. (2023). MEGA: Multilingual evaluation of generative AI. Conference on Empirical Methods in Natural Language Processing (EMNLP), 4232–4267. doi.org/10.18653/v1/2023.emnlp-main.258
  19. Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. Findings of the Association for Computational Linguistics: EMNLP 2023, 7001–7025. doi.org/10.18653/v1/2023.findings-emnlp.467

Curious what assistants really say about you?

Tell us your business and the questions your customers ask. We’ll run them through our clean-room panel and walk you through what ChatGPT, Gemini and the others say about you, and who they name instead. The first consultation is free.

Book a free consult Email us