AI Shortlists Only 3–5 Names: How to Get In

A Rotterdam team sells portable projectors into the US and Germany, and ranks second for “portable projector” at home. A buyer asks three assistants one question — backyard movie night, under $600, ships to the EU — and gets five names back, two of them outside the top ten. Theirs is absent.

The shortlist is not the top of the ranking; it is the three to five names an assistant can explain in one sentence and compare on identical fields. Comparability is the only gate a challenger can rebuild inside a quarter.

The list is assembled, not cut

A search engine answers “which pages exist.” An assistant answers “which one should I buy.” That is synthesis: fix the scene and constraints, find products that fit, attach one reason per name.

The entry ticket is not domain authority. It is fitting into one explainable sentence, manufactured from field alignment: ANSI lumens, resolution, throw distance, decibels, warranty, certification. A name whose differences cannot be stated gets merged into “other brands also offer similar models,” or dropped. List length is a compromise, not a rule: more names mean less explanation each — an opening for a small brand with solid specs, a hole for a large one that only tells stories.

The list moves with the phrasing. Budget $600 to $1,200, backyard to campsite, US to Germany. A shortlist is coverage across questions; test one phrasing and you tested one shelf.

Three gates, run in series

Before a name earns a slot, three things are checked in order. Fail the first and the rest do not matter.

Gate	What the model checks	Passing looks like	Stuck looks like	Smallest fix
Category relevance	The right product, category, and scene	Title, category, schema use the same words	Filed next door, or not identified	One category name and alias set, site-wide
Comparability	Your fields sit beside a rival’s, in the same columns	Five filter fields, one unit system, gaps stated	Adjectives and renders; specs in captions	One table with identical columns, gap column added
Trust	Someone outside your site vouches for you	Three independent sources plus a certification	First-party claims, or one press release	Minimum review, list, community footprint

Budgets cluster at gates one and three. Gate two gets filed as “a PDP job,” yet it is the layer a synthesis step is shortest of. Survivors are rarely the strongest five; they are the easiest five to compare.

Five seats, and a challenger can take two

Slots fill by function, not evenly.

• Default seat. The one or two brands a buyer recalls without reading specs; memory, not numbers.
• Spec seat. The leader on one measurable metric: brightest, lightest, longest runtime.
• Use-case seat. The only sensible answer for one job, dorm room or campsite.
• Budget seat. The one that makes spending more look unjustified in a price band.
• Local seat. Local warehouse, right certification, return window matching local habit.

The default seat is not winnable in a year; the spec seat needs a real lead, not a rounded claim; use-case and local seats sit vacant because nobody wants to publish who the product is not for. A challenger takes a use-case seat, proves it with same-unit specs, then holds a local seat with warehouse, certification, and returns. Challenging an incumbent’s default seat usually earns one line: “also offers similar products.”

Inside a marketplace, reviews, fulfillment, and price set the shortlist; a better website buys no slot.

Example: from “mentioned once” to a named slot

Example — brand and figures are invented; no real product is described.

A projector brand opens with a vision statement: ANSI lumens in a caption on the third render, battery life in an image, and no sentence on who should walk away. Across 20 phrasings (budget × scene × market) it appears in 3 answers, twice as “the cheaper option.”

Three changes: a five-field table — brightness, resolution, throw distance, noise, warranty — in the units review sites use; one body sentence, “at 700 ANSI lumens it is not bright enough for a living room with daylight”; test conditions and model mapping added to two public review lists.

Four weeks later, same 20 phrasings: 9 appearances, 4 with a specific reason (“built-in battery, one outdoor session per charge”) — Example figures. The other names held; the fourth seat changed. No ranking moved.

Four steps you can run this month

Step	Action	Owner	Passes when
1 Write the phrasings	20 questions with budget, scene, market; log names, date, region	Growth	Every phrasing is reproducible, not one screenshot
2 Copy the list	Hand-copy the public fields of the three to five names into one table; mark gaps	Content	No more than two gaps, all publicly available
3 Fill the columns	One unit system, missing values stated, one “not for” sentence	Content, category	Names hidden, a stranger still picks correctly
4 Re-run in four weeks	Same phrasings; log entries and exits by question	Growth	Three phrasings move from absent to named

Step two carries the most information and needs no tool. If you cannot copy a rival’s spec, you are not comparable, and copywriting cannot close that gap.

Where this breaks

No stable phrasing yet. Answers drift in a new category; prove demand first.

High-touch and made-to-measure. Fitting, measuring, installation: a shortlist starts the decision; it does not make it.

Revenue mostly inside a marketplace. Listings and fulfillment outrank your comparison table. Fix those.

A live trust problem. A slot amplifies whatever the reviews say; settle the negative evidence first.

Shortlist entry as the only KPI. One screenshot proves nothing; track coverage by phrasing, monthly.

This piece opens the June issue on making the shortlist; the content and semantics hub holds the rest, and the next article builds the missing columns into a comparison page structure. Where trust is the gate you keep failing, start with the four trust prerequisites; to sample without trusting one screenshot, use an AI visibility baseline.

This month: pick one category, write 20 phrasings with budget, scene, and market, and copy each AI list into a table. Fill only the missing column. Re-run in four weeks — if no name moved, the column you filled is not the field the model compares on.