In May 2025, Google brought 25 of us into a closed-door room at I/O to talk about the post-click SERP. The instruction we left with was short: create non-commoditized content.

Here’s the uncomfortable part:

For 15+ years, we never commoditized content. We commoditized the sale. We built billions of pages that took a messy human problem, targeted whatever keyword the person compressed it into, and answered with some version of “buying now is the best choice.”

We did it well for our brands, and we built a sales-first web.

Oops.

AI search is the bill for that cognitive debt finally coming due. For years, we skipped past the buyer’s real thinking and answered “buy now” instead of their actual questions. 

Paying that debt off is a link building problem. Not just hyperlinks, but the links between sources, roles, risks, and decisions.

A co-citation gap analysis maps those links, showing which sources AI search trusts for each buyer role and where your content is missing from the decision.

In this piece, we’ll show you how to run it: map the sources AI search reads and cites, find the role your content doesn’t support yet, and build the asset that closes the gap.

Moving from anchor text to anchor context

We’ve been running co-citation analysis on the link graph for 15 years at Citation Labs. 

In 2011, I published a six-step co-citation method for link builders: find the pages that curate a topic, count which sources they cite together, and reverse-engineer what made those pages worth citing.

What’s changed now is the unit of work: from focusing on anchor text to anchor context.

Anchor text told a search engine what a page was about.

Anchor context tells the model why that evidence belongs in a specific answer, for a specific role, at a specific point in the decision.

The work moves from describing the page to supporting the decision.

Instead of asking which pages mention a topic together, you’re asking which sources an AI assistant trusts when each buyer role asks about the same decision, and which role your content fails to support. 

That missing decision support is the co-citation gap.

How to run a co-citation gap analysis by hand

A co-citation gap analysis counts what AI search reads and cites across same-phase, same-problem prompts for different buyer roles.

The overlaps and absences show which decisions your content doesn’t support yet.

You don’t need software for this, just one buyer decision, the committee around it, a set of prompts, an AI tool that shows its work, and a spreadsheet.

View embedded content

1. Map the committee, and write down each role’s fear

List everyone who has to say yes before the thing you sell gets bought — the real deciders, not the org chart. 

For example, for a funded biotech that’s choosing a logo, the committee is the CEO, in-house counsel, ops, and marketing.

Next to each, write what that person is afraid of. That fear is the question they’ll bring to an assistant, so write it in first person:

• CEO: Do we look like a serious, fundable company?
• Counsel: Is this name or mark going to get us sued or forced to rebrand?
• Ops: Will this survive production (from a favicon to signage to a 96-well plate label)?
• Marketing: Will this perform (get recognized, stand out cold to investors and recruits)?

Four roles, four fears, one decision.

Note: Solo purchases work the same way. The “committee” is one buyer’s competing concerns. The unit isn’t the committee, it’s the choice point.

2. Write one prompt per role

Create one controlled prompt per role, all based on the same buying decision. Keep the scenario fixed and change only the role. That makes the role’s domain of practice the variable.

Follow these five rules:

1. One shared scenario. Keep the situation fixed; change only who’s asking.
2. First person, in the role’s voice. Model their words, worries, and perspective.
3. Put the role in a specific situation. Generic prompts get generic answers. Specific situations trigger retrieval, and retrieval is the data.
4. Bundle the role’s real concerns. Real deciders carry a cluster, not one tidy question.
5. Name no brands. Brand names pollute the citation set.

Here’s the CEO prompt from our “make a logo” example: 

“I’m the founder-CEO of a biotech that closed a Series A, briefing a design firm on a new logo. I want a point of view to show them, not a blank page. What do credible, well-funded biotech brands share visually? Which startups nailed a post-raise rebrand? Which botched it? What do investors and pharma partners read into a young brand? What should a non-designer like me use to rough out a few options?”

Then add one kill-switch prompt per role: “What’s the one thing here that, if we get it wrong, we can’t undo? What would make me say no?” 

This surfaces the hard veto role.

Run more than one phrasing. The signal is what repeats, not what appears once.

3. Capture what the assistant searches, reads, and cites

Use an assistant that exposes its sources. For each role’s run, open the activity or sources panel and copy three fields into your spreadsheet:

• Sub-queries generated
• Pages read
• Pages cited

Tag each read page by type as you go: 

• Forum/hub (Reddit, aggregators, etc.)
• Primary/official (a regulator, a standards body, etc.)
• Vendor (services and solutions)

The mix shows what kind of evidence each role trusts. 

In our example, nearly a third of everything read was Reddit, and over 40% was high-volume hubs.

Your pages will sort into three states:

• If it’s read and cited, it was consulted and used.
• If it’s read but not cited, it was consulted but dropped. That’s a content problem.
• If it’s not read, it was never consulted. That’s a discoverability problem.

The dropped-but-read pages are where the cheapest wins hide.

4. Build the citation matrix

Now, turn your sort list into a matrix — this is the co-citation analysis.

Make one row per unique cited URL, one column per role, and one count column showing how many roles cited it.

Note: One row per cited URL

Here’s what the matrix looked like in our “make a logo” example:

Sort by count. You’ll see which sources are shared, which are role-exclusive, and which roles have almost no overlap.

5. Find the role with the veto power

With the matrix sorted, look for three structures. 

The shared core is everything cited by 2+ roles. If it’s nearly empty, the committee is disjointed (you serve it seat by seat). If it’s substantial, the committee is convergent (you win the commons and the gatekeeper).

An isolated decider is a decisive role whose cited sources are mostly role-exclusive. An empty edge is two must-agree roles that share nothing.

Then set priority: whose “no” is final, and whose sources overlap least with everyone else’s?

That’s your veto × isolate seat. Build content there first.

Capture this in a spreadsheet you can update and rerun:

6. Add the phase axis

Run the analysis again, but move the same decision forward in time. Keep each role the same while changing the scenario:

Choosing → Rolling Out → Getting Value and Renewing

The cited set shifts at every stage while the owner of the answer hands off as the buyer moves.

Note where you’re present in one stage but missing in others. 

Those gaps are asset targets.

7. Create a content and outreach plan

Your matrix gives you a prioritized content strategy:

1. Gatekeeper: The veto-isolate seat, decisive and unserved.
2. Empty edges: Where two must-agree roles share nothing.
3. Shared core: Sources cited by multiple roles.
4. Phase gaps: Stages where your brand appears, then disappears, by role.

The new work is placement, and the model already told you where. The sub-queries show the domains it trusts for each role.

In our example, the CEO run ran site:businesswire.com, site:fiercebiotech.com, and site:prnewswire.com; the ops run ran site:developers.google.com/search/docs

Nearly every role appended “official” to chase primary sources. 

That’s your placement and outreach list — written by the AI assistant. Earn your evidence on the surfaces it already uses when answering that role.

And remember the unit you’re building in: anchor context, not anchor text. 

The asset’s job isn’t to show a popular option. It shows: who this helps, what it solves, when it fits, and why it belongs — for this role, at this choice point.

8. Measure the impact

Lock your prompt set and rerun it in a few weeks. (We’ve seen results in less than 3 weeks with our clients.)

Rebuild the cited column and compare it to the baseline. Look for three movements: your brand appears where it didn’t before, placed sources enter the gatekeeper’s answer, or an empty edge starts to fill.

It’s tedious by hand, but the loop is always the same: baseline, build, re-run, compare. 

The cited set is the scoreboard.

If you’d like a faster way to run this, reach out to Citation Labs. We’ll share our prompt stack with you.

How to turn the matrix into decisions

Once the matrix is sorted, review the shape of the citation set.

If a few sources are cited by most of the committee, it’s convergent. There’s shared ground to win. If the top is nearly bare, it’s disjointed. Every role is in its own world, and generic “bottom of funnel” content won’t carry the decision.

In our make-a-logo example, exactly one source was cited by multiple roles: Canva’s brand-kit page. At the read level, Reddit, Wikipedia, and arXiv showed up across roles, but almost none of that survived into what got cited. 

Now find the seat that shares almost nothing with anyone. 

That’s the gap. 

In our example run, it was Counsel: 14 cited sources, none shared with another role, all from legal, regulatory, and trademark sources.

Lowest competition on the map. Highest leverage.

You may also find an empty edge: two roles that both have to say yes but cite nothing in common. Their criteria collide with no content in between. Each empty edge is a bridge asset waiting to be built.

Don’t be surprised by who the gatekeeper is. In the committees we’ve mapped so far, the veto-isolate has consistently been compliance, security, or legal. The org chart underweights them, but the citation map doesn’t. It shows the seat that can stop the decision and has the least content support. 

That’s where you build first.

Then check the phase re-run. When you move the committee from choosing to rolling out to getting value, the citation set shifts. Most brands focus on “choosing” and ignore everything after.

Also, the decision doesn’t end at the sale. It runs through rollout, adoption, renewal, and the next internal justification.

The move that pays is to drag the late “no” upstream, so the veto lands as a redirect rather than a demolition.

For our logo committee, “gatekeeper first” became a Founder’s Preliminary Trademark Clearance Brief: a one-page brief the founder fills out before Counsel reviews a name or mark. It captures proposed assets, commercial context, preliminary checks, and specific questions for Counsel.

Watch what that single page does:

It gives the CEO something Counsel can review before the work goes too far. The veto surfaces before money gets spent. And both sides avoid the silent standoff: “Why do they keep blocking this?” versus “FFS, did they even check?”

The brief makes them check (in a form that Counsel can act on) before either of them gets angry.

That’s anchor text becoming anchor context: “here’s exactly what this decider needs, at the moment they need it, in the form that lets them say yes.”

This is the work for link builders

For 15 years, we built links at Citation Labs.

The good ones were about putting the right evidence in front of the right person at the moment they had to make a decision.

AI search didn’t end that work. It clarified it.

Build more than hyperlinks. Build the links between choice points: the places where you reduce effort, uncertainty, and the risk of being misread.

Start with the seat that can say no and that no one is writing for.

That’s the gatekeeper. That’s the gap. 

That’s the work.