# How to Choose Target Queries That Actually Measure AI Discoverability

Shoppers ask an AI what to buy, and your product is either in the shortlist or it is not. Testing which is straightforward, but the answer is only as good as the searches you pick. How to draft 30 candidates for a product, cut them to 10 across five query types, and freeze them so the before-and-after means something.

Canonical URL: https://nile.app/blog/how-to-choose-target-queries/
Author: Jiaxing Guo
Created: 2026-08-04T21:34:59.209Z
First published: 2026-08-04T23:48:35.354Z
Last modified: 2026-08-04T23:48:35.354Z
Language: en-US

## Key takeaways

- Draft about 30 candidate searches for one product, then cut them yourself to the 10 that best cover the five query types.
- Keep the queries where your product is relevant but ranks low or does not appear; cut the ones you already win, because they cannot show progress.
- Freeze the final 10 word for word and measure those same 10 afterwards. Rewording a query mid-test destroys the comparison.
- Exclude brand names, product titles, SKUs, and model numbers: they make the test trivially easy and prove nothing about discoverability.
- Cover all five query types — product category, feature or material, use case, problem or need, and compatibility or audience.
- Adobe scored retail product pages at 66% average AI content visibility in Q1 2026, the weakest page type it assessed.

Shoppers increasingly start with a question rather than a search box. They ask an assistant what
to buy, it builds a shortlist, and your product is either in that shortlist or it is not. Most
merchants have no idea which.

Finding out is not complicated, but it does require a real test rather than a vibe. Pick the
searches you want to win, record where your product ranks for them today, do the work to improve,
then measure the same searches again. The method is simple. The part that decides whether the
answer means anything is the list of searches you pick — and that is the part most people rush.

**The short version: write down about 30 candidate searches for a product, then cut them yourself
to the 10 that best cover the five kinds of query below.** Ten per product is enough to measure
honestly, and small enough that you will actually keep it current.

The failure worth guarding against is not a low score. It is a high score that means nothing. Put
your brand name and your exact product title on the list and you will rank first for both. The
result will look excellent, and you will have learned nothing about whether a shopper who has
never heard of you can find your product.

## How to build the list

You are the one choosing, cutting and freezing this list. Nobody else can do it for you, because
the judgement involved is about your product and your buyers.

1. **Draft wide.** Write about 30 candidate searches for a single product. Aim wide on purpose:
   you are going to throw two thirds of them away, and the throwing away is where the quality
   comes from.
2. **Check where you stand.** Run each candidate wherever you care about being found — an AI
   assistant, the catalogue your products sit in, your own rank tracker — and note roughly where
   your product lands.
3. **Cut to 10.** Keep the queries where your product is genuinely relevant but ranks low, or
   does not appear at all. Cut the ones where you already rank near the top: there is nothing
   left to improve, so they cannot show progress. Cut until you have 10 that spread across all
   five query types.
4. **Freeze them.** Write the final 10 down, exactly as worded, and do not touch them again.
   Measure those same 10 after you have made changes.
5. **Compare like for like.** Same wording, same place, same settings, before and after.

Step 3 is why the draft should start at 30 rather than 10. If you write exactly 10 and four of
them turn out to be queries you already win, you are down to six and a thinner test.

Step 4 is where the discipline lives, and it is the step people break. Swapping a query halfway
through, or quietly rewording one because the first version was embarrassing, destroys the
comparison — you no longer have a before and after, you have two unrelated measurements. If a
query turns out to be wrong, note it, leave it in, and fix the list for the next cycle.

Wording matters more than it looks like it should. "Insulated water bottle for hiking" and
"insulated bottle for hiking trips" are the same idea to you. They are different strings to a
retrieval system. Once you freeze the list, you are measuring whichever one you wrote down.

## Cover five kinds of query, not one

Think like a shopper who has never heard of your brand. What would they type when looking for a product like yours? A set drawn entirely from one angle measures one thing well and everything else not at all, so spread your candidates across these five types.

| Query type | What it describes | Example |
| --- | --- | --- |
| Product category | The kind of product a shopper is looking for | `smart body composition scale` |
| Feature or material | A specific quality that matters to the buyer | `full grain leather watch strap` |
| Use case | The situation the product is used in | `fitness tracker for sleep and recovery` |
| Problem or need | The problem the shopper wants to solve | `pickpocket proof travel bag` |
| Compatibility or audience | Who it is for, or what it works with | `apple watch band for large wrists` |

A strong query does five things at once. It sounds like something a real shopper would type. It is clearly relevant to your product. Every claim inside it is backed by facts on your live product page. It is broad enough that products from other stores also compete for it. And it is specific enough to describe a real shopping need.

That last pair is the tension worth sitting with. Too broad and you are competing with an entire category on a term that describes nobody's actual intent. Too narrow and you win a query nobody searches. "Water bottle" is the first mistake. "Insulated 32oz wide mouth bottle with a bike-cage-compatible base in matte sage" is the second.

## Five traps that inflate the result

Each of these produces a better-looking report and a less useful one.

**Your brand or store name.** A shopper searching your brand has already found you. Ranking for it proves nothing about discoverability.

**Exact product titles, SKUs, or model numbers.** These make the test trivially easy. `Summit-32X` has essentially one right answer and you already own it.

**Claims your product page does not support.** If a feature or benefit is not stated on your live page, a query built on it cannot be validated, and a query that cannot be validated cannot be measured honestly. This is also the most tempting trap, because the aspirational query is usually the one you most wish you ranked for.

**Near-duplicates.** Ten distinct queries measure considerably more than ten phrasings of one idea. Duplicates crowd out variety while looking like coverage.

**Queries that point at a different product in your catalog.** Only the exact target product counts in the measurement. If a query better describes your other SKU, the test reads as a failure even when discovery is working.

| Strong query | Weak query, and why |
| --- | --- |
| `insulated water bottle for hiking` | `AquaPeak Bottle` — brand names make the test too easy |
| `non slip yoga mat for hot yoga` | `Yoga Mat Pro X2 6mm` — exact titles and model numbers prove nothing |
| `wireless earbuds with long battery life` | `best earbuds that cure headaches` — a claim the product page does not support |
| `laptop stand for standing desk` | `laptop stand` + `stand for laptop` + `laptop riser stand` — near-duplicates waste slots |

## Why this is worth twenty minutes

Two things make query selection more consequential than it used to be.

The first is that AI-sourced traffic has stopped being a rounding error. Adobe Digital Insights reported traffic from AI sources to US retail sites up 393% year over year in the first quarter of 2026, and in March 2026 that traffic converted 42% better than non-AI channels — reversing a gap that ran 38% in the other direction a year earlier. Whatever else is arguable, this is now a channel with measurable behaviour rather than a forecast.

The second is that most retail pages are not especially readable to the systems doing the finding. In the same report, Adobe scored retail product pages at an average of 66% on its AI content visibility measure. That was the weakest page type assessed, below category pages at 74%, homepages at 75%, and help or FAQ pages at 79% to 82%. Between the best and worst retailers the spread ran from 82.5% down to 54.2%.

Put together, that is the argument for measuring properly. There is real movement available, the pages most likely to be weak are the exact pages a shopping query has to resolve to, and the spread between retailers is wide enough that knowing precisely where you start is worth something.

It also explains why the measurement has to be honest to be useful. Agent-facing catalogs work on relevance matching against the text a shopper supplies. Shopify's Global Catalog documentation, for example, describes agents searching by text query with results returned by relevance, and treats inferred fields such as descriptions, attributes, and technical specifications as discovery and merchandising signals. A query set built from your brand name exercises none of that machinery. A query set built the way a stranger searches exercises all of it.

## Before you freeze the list

- About 30 candidates drafted for this product, then cut to 10
- All five query types represented in the final 10
- Every query sounds like a real shopper's search
- No brand names, product titles, SKUs, or model numbers
- Every query is supported by facts on the live product page
- No two queries mean nearly the same thing
- Each query points clearly at this product, not a neighbour in the catalogue
- The final wording is written down somewhere you will not be tempted to edit it

One product at a time. Ten queries you have thought about beats fifty you have not, and a list you
keep frozen beats a list you keep improving.

## Doing this without the manual work

Everything above is a method rather than a product. It works whoever you run it with, and it is
worth doing even if you never speak to us.

What it will show most merchants is that the ceiling is not effort, it is legibility: shoppers'
assistants build shortlists from structured product context, and a page written for a 2019 search
engine often does not make the candidate set at all. That is the gap Nile closes. We build the
agentic catalogue your products need in order to be read properly, and better ranking in AI
answers falls out of that work rather than being sold as a separate service — we do not charge for
it. There is no setup fee and no retainer; we are paid a commission only when we close a sale, so
if your numbers never move, you pay nothing.

Bring the 10 queries you just chose. They make a better benchmark than anything we would pick for
you, because you already know which searches you want to win.

## Frequently asked questions

### How many target queries do I need?

Ten per product, chosen from a draft list of about 30. The surplus matters because you will cut every query you already rank well for, so a draft of exactly 10 almost always leaves you with fewer than 10 usable ones.

### Who decides which queries make the final list?

You do. Choosing, cutting and freezing the list is your judgement, because it depends on your product, your buyers, and which searches you actually want to win. Nobody outside your business can make those calls for you.

### Why can't I use my brand name as a target query?

A shopper searching your brand has already found you. Ranking first for your own name is guaranteed and measures nothing about whether someone who has never heard of you can discover your product.

### What should I do with a query I already rank near the top for?

Cut it from the list. There is almost no headroom left, so it cannot demonstrate improvement even if everything you do afterwards works. Keep it on a separate list to defend, not on the list you are measuring.

### Can I change a query once I have started?

No, and this is the rule most often broken. The whole point is that the identical wording is measured before and after; swap or reword one and you have two unrelated measurements rather than a comparison. If a query turns out to be poor, leave it in, note it, and fix the list on the next cycle.

### How do I know whether two queries are too similar?

Ask whether they describe different shopping intent or just different phrasing of the same intent. "Insulated water bottle for hiking" and "insulated bottle for hiking trips" are one idea worded twice. If both would be satisfied by the same page for the same reason, keep the stronger one and spend the slot on a different query type.

## Sources

1. [Quarterly AI Traffic Report: AI traffic grows, but retail sites lag in AI search visibility](https://business.adobe.com/blog/ai-traffic-surge-retail-sites-not-machine-readable) — Adobe Digital Insights
2. [Global Catalog MCP](https://shopify.dev/docs/agents/catalog/global-catalog) — Shopify