Which AI citation fixes come first? Ranking gaps by impact instead of length

Most teams that run a citation audit finish with the same reaction: twenty problems, no idea which one matters. The audit says your product pages lack structured data, your FAQ blocks are thin, your facts are stale, and your comparison tables are buried. All true. But fixing all of it takes a quarter, and by then the answers have moved on. The real skill is not finding citation gaps - it is ranking them.

The ranking question has a specific shape. You are not trying to make pages "better" in the abstract. You are trying to change what an assistant says about a product. That means the only fixes worth doing first are the ones attached to a question a shopper already asks, where your page is the plausible source, and where the correct fact is currently missing or wrong. A fix that satisfies all three changes an answer. A fix that satisfies none is housekeeping.

## Step 1: Separate facts from polish

Sort every finding into two buckets. Facts are claims an assistant can be wrong about: price, availability, box contents, compatibility, warranty length, shipping cutoff. Polish is everything else: nicer headings, prettier tables, longer introductions. Polish raises a page's quality score in your head; facts change the sentence the model generates. When you only have one afternoon, facts win almost every time.

A useful test: could a competitor's page state the opposite of this and be believed? If yes, it is a fact, and it is contestable.

## Step 2: Score each gap by question frequency

Take your list of product facts and match each one to the questions shoppers actually ask. Not your top keyword - the follow-up questions. "Is the battery included?" "Does it fit a 2021 model?" "How long is the return window?" These are the prompts where assistants look for a source and either find a clean sentence on your page or find one somewhere else.

Rank the gaps by how often those questions come up in your category. A single missing compatibility range on a high-volume accessory can outrank a site-wide metadata cleanup, because it touches a question asked daily.

## Step 3: Check whether your page is citable at all

There is no point perfecting a fact on a page the model cannot parse. Before you invest in wording, confirm three things: the fact appears as text (not locked in a JavaScript widget, image, or PDF), the page is indexable and not orphaned, and the fact is consistent with your structured data. A mismatch between your JSON-LD and your visible copy is worse than either one alone - it gives the assistant a reason to prefer a third-party source.

## Step 4: Fix in order of blast radius

Work top-down by how many pages a fix covers. Site-wide template fixes - putting price, availability and shipping cutoff into every product template as text and matching JSON-LD - change many answers at once. Page-level fact fixes change one answer each and come later. This ordering feels backwards because the page-level fixes are more satisfying to ship, but the template fix is the one that moves the citation rate.

## Step 5: Re-measure with the same prompts

A fix is a hypothesis. Re-run the same question set you audited with originally, and compare which pages get cited, not just whether your site appears somewhere. If a fix did not change a single answer, it was polish mislabelled as a fact. Keep the prompt set stable so the comparison means something.

## A worked example

A store selling replacement filters has fifty SKUs, each with a compatibility list buried in a compatibility-finder widget. The audit lists it as "missing compatibility data." The prioritisation is: compatibility is a high-frequency question, the widget is uncitable (not text), and the fact is contestable. So the first fix is not per-SKU copy - it is a template change that renders the compatibility list as plain text and mirrors it in structured data. That single change makes fifty pages answerable. Per-SKU enrichment of filter lifespan, a lower-frequency question, comes in the next pass.

## FAQ