A company name without spaces misses the screening list

Measured 2026-09-19 on 97 real entity names from the U.S. Consolidated Screening List. Typed with spaces, a company's short name finds its listed record 81 times out of 83. Typed with the spaces removed — ArdAlkhair, XinjiangProduction — it finds it 3 times out of 83. Half of the misses score so low that nothing is reported at all.

Screening a company against a sanctions, debarment or entity list comes down to comparing two strings. The published discussion of how that goes wrong is mostly about transliteration, typos, nicknames and word order. This is a different failure, and it is invisible from the outside: the query looks clean, the answer looks clean, and the record was never a candidate.

What was measured

For each listed entity name we generated six spellings from its own words: the whole name, its first two words, and its first three words — each one spaced, and each one with the spaces removed. Then we scored every spelling against the listed record with a production name-similarity matcher, and recorded whether it would have surfaced as a match, as a possible match, as a disclosed near miss, or as nothing at all.

The matcher is not a reimplementation written for this page. The study bundles the matcher from the product's own source, so the scores below are the scores a real query gets, at the product's own thresholds: 0.92 for a match, 0.72 for a possible match, and 0.5 as the floor below which a near miss is not reported.

The result

Compacting the whole name is harmless. Compacting a short name is not.

Every one of the 97 whole names survived having its spaces removed. An edit-distance score barely notices: the string is one deletion per space away from the original, and the similarity stays high. That result is worth stating because it is the intuition most people have, and it is correct — for the case nobody types.

People do not type ArdAlKhairInternationalCompany. They type the short name, the way they say it out loud — and the first two words of a listed name, compacted, found the record 0 times in 53. The reason is structural rather than a matter of degree: a token-based similarity compares sets of words, and a compacted string is one word that appears in neither set. What is left is an edit-distance comparison between a short string and a much longer one, and that is arithmetic the short string cannot win.

Half the misses are silent

A miss that scores just under the line is at least visible: a system that reports near misses tells you something scored 0.5-something and let you look. 38 of the 80 misses scored below that floor, where nothing is reported by anything. The worst of them:

These are not near misses that a lower threshold would catch. Moving a threshold does not reach a score of 0.15. The comparison itself has nothing to work with.

Aervik's own products have this limit

This measurement came out of our own defect, found on 2026-09-19, and the products it describes are ours. Our screening products return no match for a compacted short name exactly as measured above. We have not found a way to close the gap that does not cost more than it buys: three ways of widening the matcher were measured over 27,000-plus unrelated pairs, and the best of them recovered a small fraction of the misses while creating new false matches between genuinely different companies — pairs like KISMAYO GENERAL TRADING and LIIBAAN GENERAL TRADING. On a screening product, a confident wrong match is not an improvement over a stated gap.

So the change we made is to say it. A single-word query that finds nothing is reported as incomplete coverage rather than a clean screen, with the reason and the instruction to re-screen the spaced legal name. Shipped 2026-09-19 and verified on the served responses. A compacted name that finds nothing now returns incomplete_coverage: true with the reason attached, on every screening product we publish. The gap in recall is unchanged — we still do not find the record. What changed is that we no longer call it a clean screen.

What to do about it, whatever you are using

1. Screen the registered legal name, with its spaces, as it appears on the list. Not the way the vendor is referred to internally, and not the way it arrived in a spreadsheet column.
2. Treat a single-token query as a shape your system may not be able to answer. It is detectable before the screen runs: if the normalised name has no spaces in it, the result deserves less confidence than the same system's answer on a spaced name.
3. Ask what your screen does with a near miss. A system that reports scores below its match floor gives you the chance to catch this. A system that returns a bare "no match" cannot tell you the difference between "nothing resembled this" and "nothing could".
4. Check the shape of your list before you screen it. Our free screening name checker flags the compacted names, the names with nothing comparable and the names of only short words in a list you paste. It runs the normalizer this measurement used, entirely in your browser — nothing you paste is sent anywhere.
5. Measure it on your own list rather than trusting this page. The data and the method are below, and the corpus is public.

Limitations

• 97 names. Enough to establish the shape and far too few to pin the percentage. A larger corpus is very likely to move the numbers; we would expect the direction to hold and the magnitude to move.
• One list and one matcher. The Consolidated Screening List, and one production name-similarity implementation at its own thresholds. A different matcher moves the rates. It does not move the 38 cases scoring below 0.5, because those share no token for any token-based method to find.
• Entity records only. Person-shaped names were excluded before the measurement, as they are in every one of our screening products. Nothing here describes how a person's name behaves.
• Generated spellings, not observed ones. The six variants are how a name is plausibly typed, derived from each listed name. We have not measured what real users actually type, and that distribution is the thing that turns this rate into a risk.

Data

Every row is published — the listed name, the spelling, the score and the outcome. Recompute rather than trust us.

• results.csv
• summary.json
• README.txt

Names come from the Consolidated Screening List published by the International Trade Administration, a work of the U.S. Government in the public domain. ITA is cited as the source and does not endorse this analysis.

Related: what a screen does not show you, the Consolidated Screening List API, and the Honesty Test, where we grade our own extraction the same way.

Corrections. If a figure here is wrong we correct it on this page with the date and what changed, and say so in the changelog. We have withdrawn a published finding before when a larger sample killed it.