AI demand in 29 markets: half of it is noise

on the third of august we pulled 75,707 rows of live search demand across 29 country markets, seeded from the things businesses actually ask for when they ask for ai: the phone answered, documents read, stock counted, numbers reconciled. the plan was to find where the demand is. what the file mostly shows is that the instrument is broken, and everybody is reading it anyway.

every number below is counted from that file. the counting is ours and it is exact; the volumes themselves are the provider’s modelled estimates, which is what everyone in this trade is working from, and it is part of the point.

nearly half the volume is not the market

671 rows are about the weather. that is 0.89% of the file, and it is 43% of all the search volume in it.

add the second contaminant and the shape gets stranger. 11,359 rows are share prices — nvidia stock forecast, s&p 500 forecast, jd sports share price forecast. that is 15% of the rows and 3.2% of the volume. fifteen per cent of the file by count, almost nothing by weight. the two together are 12,030 rows, 15.9% of the file, carrying 46% of the volume, and neither is our market.

they are there because keyword tools expand from a seed, and a seed drags in whatever shares a word with it. ours was predictive analytics services: a forecasting term, so it pulled forecasting queries — weather in the united states, share prices everywhere.

sorted by volume, the top of that us export reads: weather forecast radar at 2,240,000, national weather service forecast at 2,240,000, nws weather forecast at 2,240,000. the first row that is about our subject at all is ai chat at 1,830,000, in fourth place, and it is a consumer query rather than a business one. ranks five, six and seven are weather again.

anybody who opens an export, sorts on volume and reads the first screen is looking at the national weather service.

six per cent of the volume is the same phrase written twice

1,371 groups of rows are one phrase in different spellings inside the same market. the 1,587 extra rows carry 2,260,080 searches that are counted more than once — six per cent of all the volume in the file, from two per cent of its rows.

panama is the clearest case. four rows:

one query, written four ways. those four rows add up to 1,204,000 searches — out of a panama total, across all 623 phrases we pulled there, of 1,326,320. one query is ninety-one per cent of a country’s apparent demand.

and only three of the four collapse automatically. normalise by lowercasing, dropping accents and dropping everything that is not a letter or a digit, and secretaria virtual, secretaría virtual and secretaríavirtual become the same key. secretaria vi does not — it is a truncation, and no normalisation short of fuzzy matching will catch it. a person sees one query. the deduplication sees two. that gap is the part nobody budgets for.

the query itself is office and school administration — in spain the top result for it is a regional education ministry — not the thing we sell.

the volume is in one per cent of the phrases

the top 757 phrases — one per cent of the file — hold 86% of all the volume. the other 74,950 rows share what is left.

this is the shape of every keyword file anyone will ever hand you, and it is why “total market volume” is close to a meaningless number. a market is not its total. it is the few hundred phrases somebody with a budget actually types.

the expensive clicks are the small ones

sort by price instead and the file inverts.

median click on a phrase with 10,000 searches or more: $2.49, over the 304 of 312 such phrases the provider prices. median click on a phrase with 100 or fewer: $8.13 — but only 12,672 of those 66,376 phrases carry a price at all. the coverage is 97% in one band and 19% in the other, which is itself the finding: the provider prices a phrase when advertisers bid on it, and in the long tail most of them nobody bids on. among the ones somebody does bid on, the small phrase is worth more than three times the big one.

the broad word is cheap because the person typing it has not decided anything. the narrow one is expensive because somebody with a purchase order is typing it, and the advertisers bidding against each other know exactly what that is worth.

the map is not the one people assume

the united states holds 31,563 phrases and 30,292,260 searches a month, with a median click of $14.55. spain holds 9,038 phrases and 2,724,300 searches, median click $3.36. eleven times the volume, four times the price per click.

the gulf and north africa are, on this evidence, not a search market yet — which is not the same as not a market. it means nobody is typing these phrases there, in these languages, in numbers worth planning around. qatar, kuwait, bahrain, oman, jordan, egypt, morocco, saudi arabia and the united arab emirates together come to 3,596 phrases and 71,630 searches a month. the united states alone is 423 times that.

the united kingdom is the odd one: 12,671 phrases, second only to the us — but 796,690 searches, less than a third of spain’s. a lot of distinct ways to ask, very few people asking each one.

what we did with it

we stopped using volume as the sort. deduplicate the spellings, drop the rows whose seed dragged in the wrong subject, and then read the tail by hand, which is slow and is the whole job. what survives is a few hundred phrases where somebody is describing a problem they are paying for. two of them, from a follow-up pull on the fifth of august rather than from the file above: hvac answering service at 480 searches and $147.88 a click, cloud based veterinary practice management software at 880 and $219.28.

those are not big numbers. they are the right ones.

if you are about to do this yourself

three things, in order. deduplicate on a normalised form of the phrase before you look at any total, or your biggest market will be an accent. check what your seeds dragged in, by reading the top fifty rows and asking whether each one is your subject — ours were weather. and treat a low volume with a high click price as a signal rather than a rounding error, because that is where the buyers are.

the export is not the research. the export is the raw material, and it arrives dirty.

faq

where does this data come from?

75,707 rows pulled from a commercial keyword provider on 3 august 2026, across 29 country markets, seeded from terms about ai automation, agents, chatbots and the jobs businesses actually ask for. every number above is counted from that file exactly. the volumes inside the file are the provider’s own modelled estimates, not measurements, and the article treats them as such.

why not publish the raw file?

the provider’s terms say nothing either way about redistributing raw volumes, and silence is not permission. so the findings and the arithmetic are here, and the file stays where it is until they answer in writing.

what is the single most useful finding?

that 671 rows about the weather — under one per cent of the file — carry 43% of all the volume. anyone who sorts a keyword export by volume and reads the top is reading the wrong list.

does a big search number mean a valuable market?

usually the opposite. the median click on a phrase with 10,000 or more searches a month is $2.49. on a phrase with 100 or fewer it is $8.13. the broad word is cheap because the person typing it is not buying anything.

should i still do keyword research?

yes, but not by sorting on volume. deduplicate the spellings, throw out the seeds that dragged in the wrong topic, then read the tail by hand. the work is in the reading, and there is no export that does it for you.

if you are about to plan something on numbers like these, bring me the export and i will tell you what is actually in it before you build anything on top of it. the first conversation is an hour and it is free

Created with AI assistance.