1 of 35

AEO / GEO BRIEFING

TO AI, YOU'RE AVERAGE.

How AI search decides which brands it cites, and how to become one.

VISIBLE IN

GOOGLE · CHATGPT · CLAUDE · GEMINI · PERPLEXITY · COPILOT

2 of 35

Ask an AI what the best running shoes are.

Ask any AI for the best running shoes. Then ask it again from the phone next to you, and again from a friend in another city. You'll get more or less the same short list every time. Ask for the best hotel in Hawaii and almost everyone in this room gets Four Seasons or the Ritz. Ask who reviews running shoes the best, and you get RunRepeat, over and over.��AI isn't hunting for the best answer for you. It's returning the safe one, the same few names the rest of the internet already agreed on. Every other brand in that category, however good, just isn't in the conversation.��To an AI, most brands aren't bad. They're average. Interchangeable with thirty other companies that publish the same pages, cite the same stats, and say the same things. Average never gets recommended. The AI rounds you off to the nearest safe name and moves on.��I'll show you how AI actually decides who's average and who's the answer. I’ll go through some data from benchmark studies and tests the Momentic team ran. I'll end with the some things to do about itr. It starts with a simple fact: the AI knows your brand from two completely different places.

3 of 35

AI knows your brand from two sources.

SOURCE 1 · TRAINING DATA

Locked in until the model is retrained

Influence here takes years. It depends on coverage other people wrote on websites the brand does not own.

SOURCE 2 · LIVE SEARCH

Whatever the index returns at question time

Fresh content surfaces in weeks. This is the channel a quarter of focused work can move. Lead with it.

The AI systems that your audience uses are heavily augmented with a live index.

4 of 35

Google AI Mode vs. Google AI Overviews

86%

of the time AI Overviews and AI Mode reach the same conclusion for a query.

13.7%

of the citations are the same. Winning one platform is not winning the other.

Ahrefs AI Search Benchmark, Q4 2025 – Q1 2026. 730,000 response pairs.

Every AI surface needs its own strategy.

5 of 35

Sources rotate every 2 days. The answer doesn’t.

The cited set rotates roughly every two days, and around 45% of cited sources are new each time. The meaning holds steady: consecutive overviews average 0.95 cosine similarity. Optimizing for one snapshot is a trap.

Ahrefs benchmark, 43,000 keywords over one month.

6 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

7 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

8 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

Query fan out

Best running shoes for flat feet

Stability shoes?
Stability shoes vs. motion control shoes
Brooks vs. Asics
Best stability shoes 2026

Step two, rewrite it. It does not search your phrase, and it is not just swapping in synonyms. It fans your question out into a set of smaller sub-questions it decides it needs in order to answer you well, including questions you never asked. For "best running shoes for flat feet" it fires off the obvious reformulations, best stability shoes, best shoes for overpronation. But it also generates the implicit questions too: do flat feet even need stability shoes, what's the difference between stability and motion control, what about a related problem like plantar fasciitis. And it invents comparison and recency queries on its own based on entities is initially sees as being relevant: Brooks Adrenaline versus Asics GT-2000 for flat feet, best stability shoes in 2026. Then it runs all of those at the same time and pools the results. So you are not trying to rank for the sentence someone typed. You have to show up across a whole map of sub-questions, including the ones the searcher never knew they were asking.

9 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

10 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

11 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

12 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

13 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

14 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

15 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

16 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

17 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

18 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

19 of 35

The AI search pipeline

Read the query

02

Rewrite it

03

Pull candidates

04

Clean the pool

05

Rank what's left

06

Filter eligibility

07

Pick evidence

08

Pull passages

09

Write answer

10

Cross-check

11

Score draft

12

Send with citations

01

AI + Legacy Search

AI Only

20 of 35

Top levers

→

Query fan out coverage.
Be specific with your content.
Be unique, but not unverifiable.

21 of 35

A page clears four checkpoints before quality counts.

01 · Reachable

A bot can fetch the URL

A WAF or CDN can return 403 to AI agents while audit tools still call you crawlable. Check the 499 rate first.

02 · Readable

Words live in the HTML

A client-rendered app serves an empty shell to crawlers. The words have to be in the markup, not a script.

03 · Quotable

A passage stands alone

A sectioned page gives about one citable passage per section. A wall of prose gives zero.

04 · Consistent

Claims match the graph

Outliers and freshly rewritten brand stories get dropped at the verifier, even after passing the rest.

Before your page is ever judged on quality, it has to clear four checkpoints. Fail any one and you're out before the real scoring starts. These are the most fixable things in the whole talk, so write them down.��One, reachable. The bot has to actually fetch your URL. Your SEO tools will say the page is crawlable while your firewall or Cloudflare rules block the AI agents specifically. Check your 499 rate in the access logs, the crawler giving up before your page responds. Accessibility + technical SEO equals AI readiness. Discoverability by both humans and Al depends on the same foundations

�Two, readable. The content has to exist as HTML, not get painted in by JavaScript after load. A lot of single-page apps serve a crawler an empty shell. If your words depend on scripts running, to the AI your page is blank.��Three, quotable. A passage has to stand on its own as evidence, without the rest of the page around it. Roughly one citable passage per clean section, zero out of a wall of prose. Structure is what makes you quotable, not length.��Four, consistent. Your claims have to line up with the knowledge graph and with what other sources say. The verifier cross-checks. Contradictions, outlier numbers, and brand stories you just rewrote get dropped, because the model won't repeat what it can't corroborate.

22 of 35

Three forces decide whether you get cited.

Force 1 · Freshness

89.7%

Gets you into the window

of ChatGPT's top-cited pages were updated within a year. Stale pages fall out at eligibility, before quality is scored. The grounding window is around 13 weeks.

Force 2 · Specificity

59%

Survives the verifier

of AI Overviews contain zero brands or entities. Named brands, numbers, dates, and people survive the verifier. Generic claims get stripped.

Force 3 · Structure

1,282

Beats raw length

average words on a cited page, and length barely correlates with citation. A short, well-sectioned page beats a long wall on the same ground.

Ahrefs benchmark, 174,000 cited pages.

You've cleared the checkpoints, your facts are straight. What actually gets you cited? Three forces.��The first is freshness, and it's mechanical, not cosmetic. The grounding window is about 13 weeks, the broader recency signal about a year. 89.7 percent of ChatGPT's top-cited pages were updated within the last year. Stale pages don't lose on quality, they fall out at eligibility before quality is ever scored. Freshness means updating the substance. Changing the year in your footer does nothing.��The second is specificity. Claims backed by named brands, real numbers, dates, and actual people survive the verifier. 59 percent of AI Overviews contain zero brands or entities. The average AI Overview names just 1.3 entities; AI Mode names 3.3. The model is starving for specifics.��The third is structure. Word count has almost no correlation with getting cited. The average cited page is 1,282 words but they're all over the map. From 174,000 cited pages, a short, well-sectioned article beats a long unbroken essay basically every time. Stop writing for word count. Start writing in liftable phrases.

The cited passage is often the visible tip of a much larger content system. You do not win the passage in isolation. You win it with the cluster behind it.

23 of 35

To AI, your brand is the average of everything you publish.

Distinctive content pulls that average away from the pack. Templated content drags it back into the crowd, where competitors using the same phrases and tools already are.

Momentic · Influenceable Web report.

24 of 35

Number one on Google. Nowhere in the answer.

THE #1 ORGANIC RESULT

Runner's World

Seventy years old, the most famous name in the sport, sitting at the top of the page for "best running shoes."

THE AI ANSWER ABOVE IT CITES

RunRepeat & Fleet Feet

25 of 35

ChatGPT

5×

RunRepeat. Best overall, best daily trainer, best for comfort, the spec breakdown, the reviewer summary. Every pick that took knowing something about shoes.

1×

Runner's World. The one safe, boring pick nobody argues with.

26 of 35

Same category, opposite averages.

Runner’s World

Writes about all of running

Every distance, every topic. The average lands right where everyone else is. The machine treats it as average, because it is.

RunRepeat

Buys every shoe and cuts it in half

Measures thirty-something properties on real lab instruments. Not about who's more famous. One brand is the average. The other is distinct, and distinct is what gets cited.

27 of 35

Two ways in: be the source, or be what the source picks.

BE THE SOURCE

RunRepeat, cited by name

You earn this by doing distinct work nobody else has done. It compounds, and it's yours.

BE WHAT THE SOURCE PICKS

Adidas, ranked first by RunRepeat

The best-overall shoe is in the answer because RunRepeat ranked it first and the model took its word. Borrowed. The day RunRepeat picks a different shoe, you're out.

28 of 35

When an AI Overview appears, position one loses 58% of its clicks.

Click-through at position one dropped from a forecast 0.037 to an actual 0.016. The hit runs across the whole page: position five lost 32.6%, position ten still lost 19.4%.

Ahrefs benchmark, 300,000 keywords.

29 of 35

Around 90% of your citations come from off-site.

Ranking source pages carry a median Domain Rating of 90. For "best X" queries, 43.8% of sources are third-party listicles. Top-3 placements convert to recommendations about 4× more than positions 8–10.

28.3%

CITED PAGES WITH ZERO GOOGLE VISIBILITY

51.4%

OF SOFTWARE SOURCES ARE BEST-OF LISTS

30 of 35

Report lift and correlation, not citation counts.

10.7%

of AI Overview mentions include a link.

51.6%

of Perplexity mentions do. A rising count means different things on different surfaces.

Track presence and trend. Prove the work through lift and correlation over time.

31 of 35

Your first actions

01

Technical

Audit robots, firewall, CDN, and the 499 rate. Nothing else pays off while a bot can't fetch the page.

02

Make passages liftable

A heading and a self-contained answer per section, with named numbers and sources inside it.

03

Refresh on a real schedule

News in hours, shopping in days, evergreen guides around 13 weeks.

04

Earn/buy listicle placement

Around 44% of top-of-funnel sources are best-of lists. Get into the ones your category uses.

05

YouTube

Mentions correlate with AI visibility at about 0.74, higher than any other factor measured.

06

Translate to pipeline

Citation rate × volume × value × win rate.

Here's what to actually do, Monday, in priority order.��One. Audit the live channel first: your robots file, your firewall, your CDN, and your 499 rate. If the bot can't reach you, nothing else matters.��Two. Make your passages self-contained and liftable. Break the walls into sections, each one able to stand alone as evidence.��Three. Refresh on a real schedule. News in hours, shopping daily, evergreen about every 13 weeks. Substance, not the footer year.��Four. Earn placement in the listicles that feed your category, since they're around 44 percent of top-of-funnel sources.��Five. Show up on YouTube. Mentions there correlate with citation at about 0.74.��Six. Translate everything into pipeline language: rate times volume times value times win rate. Speak in money, not citations.

32 of 35

Good content

01

Comprehensiveness

Covers the full answer space without being bloated. Breadth that maps to the fan-out graph.

02

Point of view

A defensible stance the system can attribute to a source. Not just coverage - an opinion.

03

Evidence

Claims backed by data, research, and verifiable facts. Synthesis systems need linkifiable proof.

04

Differentiation

Information gain over the consensus. Something new the system cannot get from five other sources.

05

Topical depth

Authoritative expertise on the subject. Not surface treatment - structured domain knowledge

33 of 35

Questions to ask why your content isn’t being chosen

01

Can the system access it?

02

Can the system interpret it?

03

Can it extract useful passages?

04

Do you cover the fan-out space?

05

Do you add something new?

06

Are you selected repeatedly?

34 of 35

This is a ton of work, so we made our own tools, tech, and datasets.

35 of 35

To AI, you're average.

Until you're not.