Something Inc.Schedule a free consultation
LINK BUILDING

Link Building for AI Search Is a Memory Game

Ross Simmonds says AI models are forming a permanent opinion of your brand right now. Link building for AI search is how you make sure that opinion is the one you wrote.

TTTyler TruffiManaging Partner · AUG 16, 2026 · 12 MIN READ

Every AI model already has a picture of your company. It built that picture from whatever the web said about you, it updates the picture continuously, and it did not wait for your permission to start. Link building for AI search exists to answer one question: who is currently writing that picture, and is it you?

Ross Simmonds, founder and CEO of Foundation Marketing, put a number on it at SEO Week 2026. Speaking on the New York event's third day, subtitled 'The Ecosystem,' on May 7, 2026, he told the room: 'I think that memory is one of the most important concepts in our industry for the next 48 months. This is a pivotal moment for all of us to recognize we have an opportunity.' That is not a metaphor about brand perception in the abstract sense marketers usually mean. He meant it literally. Models retain and update an internal picture of every brand every day, whether or not that brand does anything to shape it.

Here is the part that should change how a link building program gets briefed. If you do nothing, the model still forms an opinion. It just forms it from whoever else showed up: a competitor's comparison page, a three-year-old Reddit thread that got the pricing wrong, a trade publication that quoted your rival first. Silence is not neutral. It hands the pen to someone else.

I think that memory is one of the most important concepts in our industry for the next 48 months. This is a pivotal moment for all of us to recognize we have an opportunity.

Simmonds' prescription was building genuinely memorable, shareable frameworks and distinctive language that gets picked up and repeated across the web, because repetition is what gets baked into how a model represents a category. That is a content idea on its surface. Underneath, it is a link building and digital PR brief, because the mechanism he is describing — a claim getting picked up, restated, and corroborated across independent sources — is exactly what a well-run PR program has always tried to do. The difference is the target. You used to build that repetition for search engines and referral traffic. Now you build it because repetition across sources you don't control is the input a model actually uses to decide what it believes about you.

Digital PR was already training data for AI models

The uncomfortable truth for anyone who has run a links program is that most of the infrastructure for this already exists. Trade press coverage, comparison and roundup content, community threads, review platforms, reference-style pages — these are the same source types a digital PR team has always pitched. They are also, not coincidentally, the source types most likely to get trained on, retrieved, or cited by a model answering a question about your category. The audience changed. The channel list barely did.

What has to change is the pitch itself. A traditional link pitch optimizes for one placement on one high-authority domain, because that placement carries PageRank and that PageRank compounds. A pitch built for AI search optimizes for the same claim landing, independently and in different words, across several lower-friction sources — because a model doesn't experience your brand through one page's authority score. It experiences your brand through how often the same fact shows up, stated the same way, by people who are not you. One glowing feature in a top-tier outlet moves a link graph. It barely moves a model's memory if nobody else ever repeats the claim.

WHAT LINK BUILDING USED TO OPTIMIZE FORWHAT LINK BUILDING FOR AI SEARCH OPTIMIZES FORWHY IT MATTERS TO A MODEL
Referral traffic from the linkWhether the underlying claim gets repeated elsewhereModels weight repetition across independent sources, not click-through
Domain authority of the linking siteWhether the source type is one models actually retrieve fromA forum thread with real corroboration can outweigh a high-authority listicle with none
One flagship placement per quarterMany independent restatements of the same claimA single mention is a data point; five independent mentions are a pattern
Anchor text and link equityWhether the claim is quotable in isolationA sentence has to survive being lifted out of context to get repeated
Placement volumeConsistency of the framing across every placementContradicting your own claim across sources cancels the signal instead of reinforcing it

None of this replaces the fundamentals covered in the six link types AI models actually trust or the mechanics we laid out in the anatomy of an AI citation. It reframes why they work. A citation is not valuable because Google's crawler found it. It is valuable because it is one more independent instance of the same fact, and independent instances are what a model treats as corroboration instead of marketing.

48 months
the runway Ross Simmonds gave brands to shape how models remember them, SEO Week 2026, New York, May 7, 2026
May 7, 2026
date Simmonds made the memory argument on SEO Week's Day 3, 'The Ecosystem'
4
source types worth pitching the same claim to — trade press, comparison content, community threads, review platforms
0
new channels required — the shift is which claim you pitch and where, not a new tactic

This is a program, not a mood. Below is the sequence we run when a client wants to move from hoping a model represents them fairly to actually shaping it. It starts with an audit, because you cannot fix a memory you have not read.

1PHASE 1Audit what models already say about you
THE MOVES
Run the same 15 to 20 category and brand queries across ChatGPT, Claude, Gemini, and Perplexity, and record the answers verbatim
Mark each claim as accurate, outdated, or shaped by a competitor's framing rather than yours
Trace every claim back to a likely source pattern: review site, comparison roundup, forum thread, trade article, or nothing you can identify
DONE WHENA baseline of what the model currently believes, and a shortlist of who is currently feeding it that belief.
2PHASE 2Engineer one claim worth repeating
THE MOVES
Write the single sentence that states your category position in language no competitor sentence resembles
Attach a number, a name, or a mechanism to it so it reads as a fact rather than a slogan
Test it on five people outside the company — if they cannot repeat it back after one read, it is not distinctive enough yet
DONE WHENOne sentence, quotable on its own, that a journalist or forum commenter could paste without editing it first.
3PHASE 3Pitch it to the sources models actually pull from
THE MOVES
Trade press and analyst briefings, pitched as a claim with evidence attached, not a product announcement
Comparison and roundup content already ranking for your buyer's shortlist queries, offering the specific data point it is missing
Community threads on Reddit, Hacker News, and industry Slack or Discord groups where the claim answers a question people are already asking
Review platforms, where a customer restating the claim in their own words repeats it in a voice a model treats as independent
DONE WHENThe same claim, restated independently, living on at least four source types you do not control.
4PHASE 4Get the claim corroborated by people who aren't you
THE MOVES
Ask two or three customers or partners to restate the claim in their own words in a review, case study, or public post
Seed the framework through your own team's public commentary — talks, guest posts, panel appearances — attached to named individuals, not just the brand account
Identify at least one reference-style source, industry glossary, or category explainer that could plausibly cite it
DONE WHENThe claim now appears in language you did not write, from people a model does not code as 'the vendor.'
5PHASE 5Build the monitoring loop and keep feeding it
THE MOVES
Rerun the baseline queries monthly and diff the answers against the original audit
Track which pitched placements actually landed and which never moved, and stop pitching the source types that never convert
Retire or correct any claim that started distinctive and turned into commodity language once competitors copied it
DONE WHENA standing process that tells you whether your framing is winning the model's memory or losing it to whoever is louder this quarter.

Phase three is where most teams try to shortcut the work by treating it as a distribution problem — blast the claim everywhere, count the placements. That gets you visibility. It does not get you memory, and the difference between those two things is the entire subject of the next section.

Zombie content doesn't survive contact with an LLM

Wil Reynolds, founder of Seer Interactive, spoke the day before Simmonds at the same event, on the track subtitled 'The Psychology.' His warning is the guardrail this entire program needs. 'The job of marketing was never to just be seen or be visible,' he said. 'You have to turn that visibility into something believable about your brand.' He was describing what he calls zombie content: material that racks up impressions and placements while carrying no credibility, because nobody actually believes the claim it is making.

The job of marketing was never to just be seen or be visible. You have to turn that visibility into something believable about your brand.

Apply that to link building for AI search and the risk gets sharper, not softer. A model that ingests the same unsupported claim from twenty low-quality sources does not necessarily believe it more than it believed it from one source — it may simply flag the pattern as promotional noise, the machine equivalent of a reader rolling their eyes at the fifth nearly identical press release. Worse, if the claim is false or exaggerated, the correction is public and it lives in exactly the same places the original claim did. A retracted stat on a forum thread is still indexed. A comparison page that quietly walked back a claim after a competitor called it out is still in the training data with both versions.

THE CONDITION THAT MAKES THIS WORKEvery play in this piece depends on the underlying claim being true and defensible. If it isn't, repetition just gets your exaggeration corroborated by people who noticed, which is worse than never having made the claim. This only works as a link building program, not as a wishful one.

That is also why the audit phase cannot be skipped in favor of the pitch phase, no matter how tempting it is to jump straight to the exciting part. You need to know what is currently true about how your product performs, what your customers will actually corroborate unprompted, and where your genuine differentiation sits before you engineer a sentence around it. A distinctive claim that is also accurate compounds. A distinctive claim that is not accurate gets tested by the first person who tries to repeat it and finds out it doesn't hold, and that person is now telling a different story about you than the one you wrote.

How to tell if the framing is actually sticking

Most link building programs report on links: count, domain rating, referring traffic. None of those numbers tell you whether a model's picture of your brand actually moved. You need a separate, smaller measurement loop that checks the thing you are actually trying to change.

1Query the models on a schedule, not onceA single check tells you where you stand today. A monthly rerun of the same query set is what shows whether your framing is spreading, stalling, or getting overwritten by a competitor's louder claim.
2Watch for your own language coming back to youThe clearest signal a claim has stuck is hearing it repeated in a model's answer, a customer's review, or a competitor's rebuttal, in something close to your original phrasing. That is the repetition Simmonds described actually happening.
3Separate visibility metrics from believability metricsTrack placement count separately from independent corroboration count. A claim can appear in fifteen places and be believed nowhere, which is Reynolds' zombie content problem showing up in a different report.

This loop is also where the connection back to core generative engine optimization work matters, because the audit and monitoring queries you are running for link building purposes are the same instrumentation a GEO program needs anyway. Building them once and sharing them across teams is more efficient than running two separate tracking efforts that ask the model slightly different versions of the same question.

Categories where the buying cycle is long and the vendor list is crowded feel this hardest. B2B and tech SaaS companies routinely lose the framing fight to a competitor who simply pitched a sharper claim to more independent sources, not because their product was worse but because nobody was minding what the model was learning about the category while the roadmap team was heads-down. The comparison content problem we covered in comparison pages and how they win citations is the same fight from a different angle: whoever gets their framing into the third-party comparison layer first tends to keep it, because models default to whatever is already well-corroborated rather than re-deriving a category from scratch every time someone asks.

None of this is a one-quarter project. Simmonds' 48-month window is a call to start now, not a deadline you hit and walk away from, because the same forces that let you shape a model's memory let a competitor unshape it later. The programs that hold their position are the ones that treat the monitoring loop in phase five as permanent infrastructure, not a wrap-up step.

Set expectations accordingly with whoever is funding this. A links program built for referral traffic shows results in weeks. A links program built to move what a model believes about your category takes longer to show up, because it depends on independent parties choosing to repeat your framing in their own words, and you cannot force that timeline. What you can do is stack the odds: audit accurately, engineer one true and distinctive claim at a time, pitch it to the sources that get retrieved, and keep checking whether it actually landed.

DO THIS NEXTThis week: run your 15 to 20 baseline queries across the major models and write down, verbatim, what they currently say about your category and your brand. This month: draft one distinctive, accurate claim and pitch it to one source in each of the four channels — trade press, comparison content, community threads, review platforms. This quarter: rerun the baseline queries and check whether your language, or a competitor's, is the one coming back.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

TT
Tyler TruffiMANAGING PARTNER, SOMETHING INC.

Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.

Free consultation

Let us be the last SEO agency you ever work with

A 30 minute call and a free audit of your SEO and GEO position. You keep the findings either way.