Can Apple's search suggestions measure keyword demand? 5,493 lookups against Apple's own numbers

Apple's search suggestions can sort keywords by demand, but they cannot size them. We walked 5,493 lookups against Apple's own popularity charts, measured 2026-10-26. A calibrated model ordered charted terms at Spearman 0.68 while missing their popularity by about 12 points. Where you can read Apple Ads popularity, it beats any suggestion-based estimate. In the deep tail, nothing we built measured demand.

What suggestion data can and cannot tell you

Autocomplete is the only free window into App Store demand, and it is a narrow one. Among popular terms almost everything is suggested after two or three letters, so the data runs out of room to separate them. The order of the list does carry demand, which is enough to rank phrases. It is not enough to put a number on one. Our free popularity estimator is built on this and says so on the page.

Three findings decide what the method is good for. First, suggestions saturate in the head: 451 of 468 terms on Apple's chart were suggested at all, and the median shortest prefix that produced them was two to three characters whether the term scored 41 or 94. Second, the list is ordered by demand often enough to be useful — in 4,807 pairs of suggestions that both sat on Apple's chart, the one listed higher was the more popular one 85% of the time, with a median miss of 4 points. That lets a term whose popularity you already know bound an unknown neighbor from above or below. Third, nothing survives the drop below Apple's chart. In the deep tail our model's order was close to random, even though Apple still suggested 84% of those phrases.

What we compared against

Two instruments. Apple's own popularity charts — the Apple Ads search-term popularity lists of weekly top terms per country and genre, published to Apple Ads accounts — gave us ground truth in the head: 12,519 term-country cells over 21 weeks and 5 genres, read through our own account. Astro, a paid Mac ASO app that reads the same Apple Ads popularity, gave us a second reading that reaches below the chart.

Apple's charts have a hard edge. Charted values in our pull ran from 41 to 94, and a chart that returns fewer than its maximum of 500 terms stops at a volume floor — about 44 to 46 in the United Kingdom, Germany and Spain, about 54 in Sweden. So Apple's published chart measures the head and only the head, roughly 44 and up. Everything quieter than that is invisible to it, which is exactly the range app makers care about when they are looking for a phrase they can actually rank for. The scale itself is explained in our guide to Apple's popularity score.

Astro tracks Apple closely but not exactly. Measured fresh against Apple's chart on 459 cells, Astro read Spearman 0.90 with a mean absolute error of 4.5 points, sitting a median 3 points above Apple. In August 2026 that gap was 8 points. We therefore treat an Astro reading as Apple's number plus about 3, and we say so wherever we use it.

How we walked the suggestions

We sampled 2,934 phrase-country cells across five storefronts — the United States, the United Kingdom, Germany, Spain and Sweden — in four tiers: terms currently on Apple's chart, terms that had just dropped off it, terms our apps track, and terms Astro rates above its floor. Then we walked Apple's suggestion service for each phrase the way VisualASO does and recorded two things: how little typing it took to surface the phrase, and where it landed in the list.

The walk types short prefixes of the phrase in order — the first three letters, then each whole word, then a word plus the next letter, then the full phrase — and stops at the first prefix whose suggestion list contains the whole phrase. We note that prefix length and the phrase's position. Two cheap numbers, both of which should fall as demand rises: a popular phrase needs less typing and sits nearer the top.

We walked 1,875 cells with 5,493 calls to Apple's suggestion service before Apple slowed us down. A quarter of the phrases were held out, split by phrase rather than by cell, so no model saw a storefront's copy of a phrase it had already been fitted on.

If you are building on the same endpoint, budget for its limits. Apple answered "too many requests" (HTTP 429) after about 1,700 calls in 15 minutes at roughly two calls a second from one address. Later it stopped answering altogether — 30-second timeouts — after about 5,500 calls over 3.5 hours, even when we had dropped to one call every two or three seconds. The second wall is the one that ends a long run: slowing down did not avoid it.

In the head, suggestions saturate

Among charted terms, the walk barely varies. Almost every term is suggested, and the median shortest prefix is two to three characters in every popularity band from 41 to 94. The measurement has no room left to tell them apart, so the prefix length carries almost no information about size in the head.

Head of demand: 468 cells on Apple's chart (week of 2026-10-11), walked 2026-10-26	Result
Suggested at all	451 of 468
Median shortest prefix that offered the term	2–3 characters, in every band from 41 to 94
Higher-listed of two charted suggestions is the more popular	85% of 4,807 pairs (median miss 4 points)
Raw suggestion score vs Apple's chart (Spearman)	0.54
Best single walk feature (Spearman)	0.63
Astro on the same cells (Spearman)	0.90

List position is the part that still works. Across 4,807 pairs where both suggestions sat on Apple's chart, the higher-listed one was the more popular 85% of the time (median miss 4 points). This is a comparison, not a measurement, and it is best used as one: if you know a term's real popularity from Apple Ads, every term listed above it is probably more popular and every term below it probably less. That bounding trick is worth more in practice than any single estimated number.

A raw score built from the walk alone agreed with Apple's chart at Spearman 0.54 across all cells. On the same cells, Astro agreed at 0.90. The free data is in a different class from the paid reading, and no amount of fitting closed that distance.

A calibrated model, and the error it carries

We fitted a linear model on the walk features plus the ratings counts of the top ten apps for each phrase. On held-out chart cells it reached Spearman 0.68 with a mean absolute error of 11.7 points. When a stored Apple chart supplied anchor terms, it improved to 0.69 and 10.8. Our earlier simple mapping read 0.58 and 13.0. Astro, read fresh and shifted 3 points down, read 0.83 and 4.1.

Held-out cells	n	Old simple mapping	Model, with chart anchors	Model, no anchors (the free tool)	Astro read fresh, −3
On Apple's chart	129	0.58 · 13.0	0.69 · 10.8	0.68 · 11.7	0.83 · 4.1 (112 cells)
Astro above its floor	205	0.57 · 16.1	0.71 · 12.1	0.66 · 13.5	—
… off Apple's chart only	186	0.54 · 16.8	0.66 · 12.3	0.62 · 14.0	—

Each cell: Spearman correlation · mean absolute error in points on Apple's 1–100 scale. A constant guess has a mean error of 11.4 on the charted cells and 23.5 on the off-chart cells.

The uncomfortable number is this one: on charted cells, guessing a single constant for every cell gives a mean error of 11.4 points, about the same as the model's 11.7. That is not because the model is worthless. It is because charted terms all live between 41 and 94, so a constant is already close. What the constant cannot do is sort. The model's value is order, and the honest way to describe it is as a ranking instrument with roughly twelve points of slop on any individual reading.

By storefront, the pattern is consistent: English and German storefronts behave best, Spain worse, Sweden worst and on too few cells to argue about.

Storefront	Raw suggestion score (Spearman)	Calibrated model, held-out chart cells
United States	0.54	0.73
United Kingdom	0.60	0.74
Germany	0.55	0.77
Spain	0.35	0.62
Sweden	0.28	0.56 (12 cells)

The raw score is read on all charted cells; the model on held-out charted cells, without anchors.

The tail: where nothing measures

Below Apple's chart we had to use Astro's reading, shifted 3 points down, as the target. Across 186 such cells the model read Spearman 0.62 with an error of 14.0 points. The per-band numbers show how much of that agreement comes from telling the bands apart rather than ordering terms inside one:

Astro reading	Model agreement (Spearman)
41 and up, off Apple's chart	0.39
21–40	0.28
6–20	0.16

In the 6–20 band — the quiet phrases an indie app can realistically own — no model we built measured demand. Suggestions still offered 84% of those phrases, which is the trap: Apple autocompletes terms few people search for, so presence in the list proves nothing about size.

One thing does survive. The model separates "some demand" from "none measured". On cells Astro reads at its floor of 5, the model returned a number for 90% of them, with a median of 12 against a median of 31 for cells Astro actually measures, and told floored from measured apart with an AUC of 0.84. A coarse yes-or-no about whether anyone searches a phrase is reachable from free data. A number is not.

What this means for VisualASO's demand reading

VisualASO prefers the best instrument it can reach for each phrase, in this order: Apple Ads popularity if you sign in, then the calibrated suggestion estimate, then an Astro import (shown 3 points lower and labelled as such), then a ratings proxy as the weakest fallback. Every number carries a letter naming where it came from, so you can see which instrument you are reading. Signing in is covered in our Apple Ads notes, and every reading's source is listed in the numbers reference.

A floor value is shown as "?", never as zero. Those are two different claims — "nobody searches this" and "no instrument on hand can see this" — and only the second one is ever true of a floored cell. Writing zero there would invite you to drop a phrase that might be worth keeping.

The free estimator runs the same model without chart anchors, so its accuracy is the "without anchors" row above: about 12 points of error on terms in the head, and no ability to size the deep tail. In that range it prints "Low" instead of a number. On our measurements, a confident two-digit figure for a long, quiet phrase cannot come from suggestion data alone.

Limits

Five storefronts and five genres of Apple's charts, with Sweden's sample too small to carry its own conclusion. Below the chart our target was Astro's reading shifted 3 points down, and Astro's offset against Apple drifts — 8 points in August 2026, 3 points in October — so tail errors inherit that drift on top of their own.

Apple's popularity data went through a disruption during the period we measured. aso.dev reported that from 29 September 2025 the number of US terms rated above 5 fell by 77%, and on 28–29 October 2026 reported the count rising again from 23 October 2026. Our charts are from August to October 2026, inside that window.

The result tables are available as a summary CSV; it holds the aggregate results, not term-level data. Cite as: VisualASO, "Can Apple's search suggestions measure keyword demand?", measured 2026-10-26, visualaso.com/research/app-store-search-suggestions-demand.

Questions

Can I get App Store keyword search volume for free?

You can get a rough ranking for free, not a volume. Our calibrated model missed Apple's own popularity by about 12 points on charted terms and could not size phrases below Apple's chart at all, where many keyword decisions happen.

Is Apple Ads popularity better than an autocomplete estimate?

Yes, by a wide margin. On the same cells, Astro's Apple Ads reading agreed with Apple's chart at Spearman 0.90 against 0.54 for a raw suggestion score, so if you can sign in to Apple Ads, read popularity there instead.

If Apple suggests my keyword, does that mean it has demand?

No. Apple suggested 84% of the phrases in the quietest band we tested, where Astro reads popularity between 6 and 20. Appearing in the suggestion list tells you the phrase exists as a search, not that it is searched often.

Why does VisualASO show "?" instead of a popularity number?

Because the instrument in use cannot see that phrase. A floored reading means "not measured", and showing zero would read as "no demand", which we have no basis to claim.

Why is the accuracy worse in Spain and Sweden?

We do not know the cause, only the result: the calibrated model reached 0.62 in Spain and 0.56 in Sweden on held-out chart cells, against 0.73 to 0.77 in the United States, the United Kingdom and Germany. Sweden's figure rests on 12 cells.