What is an anchor question?

A benchmark question that is never edited or retired for as long as its category is measured, so the identical wording can be asked year after year and the answers compared.

Why not just update questions as the market changes?

Changing a question changes what is measured. Updated questions cannot be compared with older answers, so anchors hold part of the set still while the rest evolves.

How many anchors does a set need?

Citewake's published method makes 20 of its 60 scored questions per category anchors, about seven in each tier.

When can an anchor be retired?

Only when it stops representing the category, with a dated reason, a method version and approval. A replacement runs alongside it for at least three editions first.

Anchor questions: comparing AI answers across years: Citewake

An anchor question is a benchmark question that never changes. Because its wording is fixed, the same question can be asked in 2026, 2028 and 2030 and the answers compared directly. Anchors give a measurement set the one thing it otherwise lacks: a stable point to measure from.

What problem do anchors solve?

Every benchmark faces a tension. Markets change, so questions need to change. But if questions change, results can't be compared over time.

Citewake's published method handles part of this by versioning: changing any question makes a new benchmark version, and numbers are compared within a version only. That keeps comparisons honest, but it also means a long run of editions is cut at each version change.

Anchors fill the gap. They carry across versions unchanged, so something is always comparable.

How do anchors work in the method?

The published method sets out these rules.

A share of the set. In each category, 20 of the 60 scored questions are anchors, about seven per tier.
Never edited or retired in the ordinary course. An anchor's text is immutable within a benchmark version and carries into the next version unchanged by default.
Asked every edition. They run for as long as the category is measured.
Retired only with cause. An anchor may be retired when it stops representing the category, for example when a feature it names becomes universal or disappears. Retirement needs a dated reason, a method version and approval.
Replacements overlap. A replacement runs beside the retiring anchor for at least three editions before the old one stops, so the series can be bridged. The retired anchor's history stays published.

What does this let you see?

The same question, word for word, across years. For a buyer or an analyst that shows things a changing set cannot:

which brands arrived in a category's answers and which left;
which sources began to be cited and which disappeared;
how answers grew longer or more cautious;
which AI named a new brand first;
what shifted when a provider changed its model.

The method calls this the archive: every answer kept, every month, so anyone can pick a question and see what each AI said in each past edition, beside what it says now. Why every AI answer should be kept covers the storage side of that.

What makes a good anchor?

An anchor has to last. A question that depends on a short-lived trend, a feature in the news or a particular product is a poor anchor, because it will stop making sense.

Good anchors tend to be:

Durable in wording. They describe a need that will exist in years, such as choosing between kinds of tools, a typical price question or a how-to that doesn't depend on a product.
Broad across tiers. Some in the head, some in the mid and some in the tail, so that the series reflects the whole category.
Free of brand names. Naming a brand ties the question to a company's fortunes.
Clear in meaning. Anyone reading it in five years should understand the need without extra context.

What are the limits of anchors?

Be honest about three.

The AIs change under them. A fixed question asked of a changing model shows the model's change as much as the market's. That is why the method marks provider model changes on every timeline and labels comparisons across them.
They are a sample. Twenty questions can't cover a whole category. They show the direction of travel, not every detail.
A fixed question can age. A question that was typical when it was written may become odd. The retirement rule and the overlap period exist for that reason.

Can you use anchors yourself?

Yes. If you track AI answers for your own business:

Pick five to ten questions that will still matter in two years.
Write them once, with no brand names, and never edit them.
Ask them on a regular schedule, across the assistants you care about.
Add other questions around them as your interests change.

Keep the anchors in their own list, with the date you started. When you look at results a year from now, you can compare like with like. How to build a benchmark set of questions covers the wider set, and how to track AI visibility over time covers the routine.

What is the difference between an anchor and a trend line?

A trend line is a statistic. An anchor is a question. The trend is only as good as the question under it. If the question quietly changes, the line breaks without anyone noticing. By keeping the question fixed and the full answer stored, anchors make it possible to see the line and also to read what lies behind it.

How would you check your own answers against an anchor?

Ask the same anchor question in a fresh, signed-out session in each assistant, save the full text and note the date. A quick way to see how assistants treat your own site today is the Citewake checker, though a serious series needs the same questions repeated on a schedule.

Where is the full method?

On the methodology page. It describes anchors as part of a method. The first published edition is due in early 2027, and the archive starts with it. Earlier years aren't back-filled with anything that wasn't collected at the time.