We pulled 980 Singapore-relevant keywords about SEO services and checked what Google actually returns for each one. Keywords carrying 68% of the total search volume now trigger an AI Overview — an AI-written answer sitting above the ranked results. In our own category, two-thirds of the demand meets a machine-written summary before anyone reaches a link.
We are an SEO agency publishing a number that makes SEO look harder to sell. It seemed better to measure it than to wait for a client to ask.
What exactly did we measure?
In July 2026 we exported broad-match keyword data for the phrase “SEO service” in the Singapore market. The raw export held 1,167 keywords. We removed every keyword aimed at another country — searches for agencies in Manchester, Bali, and a long tail of American cities — which took out 187 rows. That left 980 keywords carrying 12,630 searches a month.
For each of those we read the SERP features recorded against it: whether Google shows an AI Overview, a Local Pack, ads, and so on. Then we weighted by search volume rather than counting keywords, which turns out to matter enormously.
What did the numbers actually show?
| What we checked | Keywords | Monthly volume | Share of volume |
|---|---|---|---|
| Returns a Google AI Overview | 22 | 8,600 | 68% |
| Returns a Local Pack (map results) | 19 | 8,500 | 67% |
| Returns both | 12 | — | — |
| Everything else | 958 | 4,030 | 32% |
Twenty-two keywords out of 980. Two percent of the rows, and more than two-thirds of the demand. If you counted keywords instead of searches you would conclude that AI Overviews are a rounding error in this category, and you would be wrong by a factor of thirty.
Why do 22 keywords carry two-thirds of the demand?
Because search volume in a commercial category is not spread evenly — it is piled up on a handful of obvious phrases. The single biggest keyword in our export carries 3,600 searches a month on its own, which is 28% of the entire Singapore-relevant universe from one phrase. It is a search for local SEO services, and it returns both an AI Overview and a Local Pack. A business owner typing it sees a written answer, then a map, and only then the ranked links.
The four largest keywords between them account for more than half of all the demand we measured. Everything else is a very long tail of small variations.
The part of the export that was worth nothing
Here is the finding that surprised us most, and it is not flattering to the way our industry uses these tools. Of the 980 Singapore-relevant keywords, 951 had no intent classification, a keyword difficulty of zero and a cost-per-click of zero — meaning the tool had no meaningful data on them at all. Those 951 rows carried 2,620 searches between them, an average of under three a month each.
Ninety-seven percent of the rows in a keyword export were, for planning purposes, noise. Only 23 keywords in the entire file carried 50 or more monthly searches, and those 23 accounted for 77% of the volume.
This matters if you have ever been shown a content proposal built on a keyword count. A plan that promises to target a thousand keywords, or to publish a hundred articles because a hundred rows appeared in a spreadsheet, is usually a plan built on the noise. The honest version of that proposal is much smaller and much more specific, and it is harder to sell because it looks like less work.
What does this change about SEO?
Less than the headline suggests, and more than most retainers have adjusted for.
It does not mean search traffic disappears. The keywords with AI Overviews still carry the highest commercial value in the file — the two most expensive keywords by advertising cost both sit in that group, which is the ad market’s way of saying those clicks still convert. Buyers are still arriving. They are arriving through a different door.
What it does mean is that the target has split in two. A page can now win by ranking, or by being the source the AI summary quotes, and increasingly it needs to do both. The useful discovery is how much the underlying work overlaps: a site that can be crawled cleanly, content structured so a specific question receives a specific answer, unambiguous signals about what your business is, and real authority behind the claims. That is the same list either way.
Which is why we treat the AI-answer layer as part of doing SEO properly rather than as a separate service with its own invoice. The work was always going to be the same work.
What should you actually do about it?
- Check your own keywords, not ours. The 68% is specific to searches about SEO services. Your category has its own number and it is not hard to find — search your top ten terms and look at what sits above the results.
- Weight by volume when you plan. A keyword list sorted alphabetically hides where the demand actually is. Sorted by volume, most lists collapse to a handful of terms that matter and a tail that does not.
- Look at what the AI answer currently says about your category. It is naming someone. Finding out who, and on what basis, is a twenty-minute exercise that most businesses have never done.
- Stop paying for position alone. If your reporting shows only rankings, it is describing a surface that two-thirds of your demand now passes over on the way to something else.
Two percent of the keywords. Sixty-eight percent of the searches. If you plan by counting rows, you will miss where all the demand actually is.
Methodology, so you can check this
A number this convenient should be reproducible. Here is exactly how it was produced.
| Parameter | Value |
|---|---|
| Tool | Semrush Keyword Magic Tool |
| Match type | Broad match |
| Seed term | “seo service” |
| Database | Singapore |
| Collected | 29 July 2026 |
| Raw export | 1,167 keywords |
| Exclusion rule | Keywords naming a non-Singapore location removed (187 rows) |
| Analysed universe | 980 keywords / 12,630 monthly searches |
| AI Overview flag | Semrush’s recorded SERP-feature field, not a live manual check |
| Calculation | Sum of volume for AI-Overview keywords ÷ total volume of the analysed universe |
Download the 22 AI-Overview keywords (CSV) — keyword, monthly volume, difficulty, CPC, intent and recorded SERP features.
What this measurement does not establish
- It is a recorded feature flag, not a live observation. Semrush records whether an AI Overview was seen for a keyword. We did not re-run all 980 searches manually, and AI Overviews are not shown uniformly — Google varies them by phrasing, location, device and history.
- Search volumes are modelled estimates, as they are in every keyword tool. Treat the proportion as sound and the absolute numbers as approximate.
- Branded and near-duplicate keywords were not removed beyond the geographic exclusion. The universe is the export as supplied, minus foreign-location terms.
- One category, one market, one date. SERP features change. We intend to re-run this and publish the movement; if you want the comparison when it exists, ask.
The limits of this data
Three caveats, because a number this convenient deserves them. This is one tool’s view of one category in one market at one point in time; SERP features change and the figure will move. AI Overviews do not appear uniformly for every user — Google varies them by query phrasing, location and history, so a recorded feature is a strong indicator rather than a guarantee. And search volume figures from any keyword tool are modelled estimates, not counts.
None of that changes the shape of the finding. The demand in this category is concentrated in a small number of phrases, and those phrases are the ones Google has chosen to answer itself.
Frequently asked questions
What is a Google AI Overview?
An AI Overview is a summary Google writes itself and places above the ranked results. It answers the query directly, drawing on several sources, and names some of them. The user gets an answer without visiting a website, though the cited sources still receive some traffic.
Does an AI Overview mean nobody clicks through?
No, but it changes who clicks and why. Someone who wanted a quick fact is satisfied by the summary and stops. Someone who wanted to hire a vendor still clicks, but often clicks what the summary named. Being cited in the answer becomes a route to the click rather than a replacement for it.
How did you measure this?
We exported broad-match keyword data around SEO services for the Singapore market from Semrush in July 2026, removed rows targeting other countries, and read the SERP features Semrush records for each remaining keyword. We then weighted by monthly search volume rather than counting keywords, because the volume distribution is extremely uneven.
Why weight by volume instead of counting keywords?
Because counting keywords would badly mislead. Only 22 of 980 keywords carry an AI Overview, which sounds negligible at 2% of rows. Those 22 keywords carry 68% of the actual searches. Keyword counts flatter small terms; volume weighting reflects what people actually type.
Is this specific to SEO keywords, or does it apply to my industry?
The 68% figure is specific to searches about SEO services in Singapore. The pattern is not. Commercial research queries across many categories are increasingly answered above the fold. The only way to know your own number is to check your own keywords.
Does this mean I should stop investing in SEO?
The opposite, but you should change what you are buying. The work that earns an AI citation is largely the work that earns a good ranking: clean crawlability, clear structure, direct answers, and genuine authority. What loses value is the part of a retainer that only ever chased position on a page fewer people now read to the bottom of.
Where this leaves your SEO
If you take one thing from this: the question is no longer whether to do SEO, it is whether the SEO you are paying for is building anything an answer engine can use. Most of the work is identical. The reporting, the structure and the priorities are not.
If you want to know what the AI answers in your category currently say, and whether your site is anywhere in them, we will check and tell you what we find. See how we build for both surfaces, message us on WhatsApp, or send an enquiry.
Sources
Written by Charlotte Zhang, Operations Partner at Kaizenaire Pte Ltd (UEN 201932071D). Charlotte co-founded Kaizenaire with Ken Tan in 2019 and runs client delivery, after four years in operations at OCBC and HSBC.