We asked ChatGPT to name Singapore's best baby sleep consultants. Our client wasn't on the list.

We ran one query across ChatGPT, Claude and Gemini: "Who are the top baby sleep consultants in Singapore?"

Two of the three did not mention our client. The third put her third and quoted a package she had stopped selling, at a price roughly SGD 590 below what she actually charged.

This is a business with more than a hundred five-star Google reviews, a decade of practice between its consultants, and a website that performs perfectly well in ordinary search. None of that helped, because none of it was written in a form the software could use.

Because a normal audit looks at the page a human sees. Speed, mobile layout, meta descriptions, broken links. Every one of those was fine here.

A widget looks perfect to a parent and is blank to the software deciding whether to recommend you.

Her reviews sat inside a JavaScript widget. Her booking sat inside an embedded frame. Both render beautifully in a browser and return nothing to the crawlers that matter, because GPTBot, ClaudeBot and PerplexityBot mostly fetch raw HTML and do not run JavaScript.

So the most persuasive things about the business, the reviews and the ability to book, were exactly the things no assistant could see.

We made the site legible. Wrong addresses now return a real error page. We published a machine index of all 47 pages, an llms.txt summary of the business, and explicit permissions for 20 AI crawlers.

We removed what was false. The dead email, the stale prices, a misspelled city sitting in the code behind three separate pages, and the three conflicting sets of opening hours.

We declared what the business is. One entity, one identity, linked to its social profiles, serving a stated age range in a stated set of countries. The site had been describing itself as two unrelated companies.

We made booking machine-readable. A free 20-minute call is now published as an action, at a named URL, with its availability stated. The difference between "they offer consultations" and "you can book here" is the difference between a mention and a referral.

We published the commercial terms. Ten questions answered in plain HTML: payment, surcharges, twins pricing, rescheduling, refunds, ages, and a stated WhatsApp response time. No competitor in the category publishes a response time.

Even after the site was fixed, one assistant kept quoting prices that were wrong. Not from her website. From Singapore parenting roundup articles published years earlier.

Much of what an AI says about your business is not on your website at all.

That matches the wider pattern. Eight of the ten most-cited URLs across AI platforms are "best of" listicles, and blog content is the single most-cited page type in AI Overviews. Freshness compounds it: pages updated within 60 days are around 1.9 times more likely to appear in an AI answer than pages left alone.

Fixing your own site is necessary and not sufficient. The other half of the job is being present, and current, in the places assistants read about your market.

Check what your site says when nobody is looking. Structured data, error pages and crawler rules are invisible in a browser and are most of what an assistant reads. Ours had a misspelled city in it for years.

Anything inside a widget does not exist. Reviews, booking, testimonials, pricing tables. If it matters commercially, it needs to exist as text as well.

Publish the awkward answers. Refunds, surcharges, response times, what happens when plans change. Assistants recommend the business whose terms they can state confidently, and most of your competitors have published none of theirs.