Seven locale mistakes that keep a Singapore business out of AI answers, starting with the one we made ourselves

Seven mistakes account for most of it, and they are ranked here by what each one costs rather than by how technical it sounds. The expensive ones are not the obvious ones. The single most expensive locale defect we have measured anywhere was a page whose prices existed in several currencies and served only one of them to anything that was not a browser, and it was our own.

The seven, in the order worth fixing

1. Currency figures that are not in the served document. If your prices are converted by a script, a crawler sends none of the signals a browser sends and gets whatever your fallback is. Put every currency you sell in into the HTML on every request.
2. No country stated anywhere a machine can read. A Singapore address, a Singapore phone format and the word Singapore in the page text. Obvious, and frequently missing from the pages that need it most.
3. GST silence on a price. A figure with no indication of whether tax is included invites a reader to assume the cheaper reading, and for a Singapore buyer the 9 per cent matters. State it beside the number.
4. A service area that does not match reality. Pages implying regional coverage you do not have, or Singapore only coverage when you serve the region. Either one produces enquiries you cannot fulfil or misses ones you can.
5. One page trying to serve several markets. A page addressing Singapore, Malaysia and Indonesia at once addresses none of them specifically enough to be quoted. Split it, and settle the architecture first.
6. Dates and formats that are ambiguous. Write the month in words. A date that could be read two ways on a page about a deadline is a support ticket waiting to happen.
7. Hreflang that contradicts the content. A tag claiming a Singapore page while the copy prices in another currency and names another country's regulator is a signal fighting the page it is on.

The one we got wrong, with the figures

Our own pricing page is the clearest example of mistake one, and we publish it because the cause is more useful than the fix. Fetched as a crawler, the page returned HTTP 200 and 88KB of which 55KB was script payload, and the visible first response carried rupee figures and no dollar, pound or euro figure at all, because the conversion ran in the browser.

The detector was never broken. A browser sends a country header, a time zone cookie or an accept language header and gets its own ladder. A crawler sends none of the five signals, so the detector reached its last line, which reads we could not tell, so rupees. That default is deliberate and protects an Indian buyer whose phone ships with a British locale from being shown a ladder more than twice his own. The defect was that the other currencies were never in the HTML at all. It was fixed the same day with a server rendered chapter that prints every plan in every region on every request, and the page now carries 15 dollar figures, 5 pound, 5 euro and 31 rupee in its visible text.

And our own gate could not see it. We already run a check that fetches a page as each AI crawler and as a browser and fails when a figure a browser sees is missing for a crawler. Run from our own network, both got rupees, so there was no difference to find. A comparison between two requests that share a blind spot cannot reveal the blind spot, and the replacement asserts the arithmetic instead: at least three distinct currencies in the table, with the right one for each named market.

The residual gap is worth stating because it applies directly to a Singapore reader. Our pricing page offers six currencies in its selector and only the rupee and dollar figures are in the served HTML, and the machine readable llms.txt file carries those same two. There is no Singapore dollar figure in either, so Singapore takes the dollar column: read on 22 October 2026, Growth is $199 a month outside the United States and the United Kingdom for 100 buyer questions across all five engines, and SEO Start at $99 reads zero engines.

What a locale mistake costs in an answer

On 22 October 2026 we put 33 batches of buying questions to ChatGPT and Gemini with Singapore written into the text of each one, because the search interface we use accepts country, language and location parameters and ignores all three. The snapshot taken that evening holds 64 blind readings, 32 per engine. The engines returned 757 distinct hosts and reached 39 distinct Singapore government hosts 90 times across 19 readings.

In 26 of the 64 readings they named no business at all. A locale defect does not usually produce a wrong answer about you; it produces no answer about you, because a page that cannot be read as a Singapore page is not a candidate for a Singapore question. That is why the ranking above puts machine readability above tagging.

Questions about locale problems on a Singapore site

How do I tell whether my site is being shown to the right country

Fetch it yourself with a plain request that sends no cookies, no accept language header and no user agent, and read what comes back. Whatever that response contains is what a machine with no signals gets, and that is the worst case your page has to survive.

Does a currency selector count as having the price on the page

Only if the figures are in the document. A selector that converts in the browser means the figures exist for a person and not for a crawler, which is exactly the defect described above.

Should I write Singapore into every page

Into every page where the market changes the answer, which is most commercial pages. A generic page serving everybody is quotable for nobody.

Is hreflang worth setting up for one market

For a single market site it buys very little. It earns its keep once you have separate pages for Singapore and at least one other market, and our guide to running one programme across several markets from Singapore covers the architecture decision that has to come first.

What is the quickest win on this list

The GST line beside your prices, and the Singapore address in your page text. Both are an afternoon and both remove an ambiguity a reader notices.

The thing worth checking after you fix them

Every item on this list is checkable by a person with a terminal and half a day, and none of them tells you whether the fix changed an answer. That needs a reading before and a reading after, on the same questions, with both dates recorded. We took ours for Singapore on 22 October 2026 and our own business was named zero times and cited zero times across all 64 readings, which is the baseline a locale fix has to move. AI Knows Us keeps that question list running across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and reports reached and named as two separate figures, because a locale defect breaks the first and a content problem breaks the second.