SEO After AI

AI answers broke the guarantee that ranking produces visits. This note documents how Exiid rebuilds organic acquisition around citations, entities, and surfaces beyond the SERP — the working model our Growth engine runs today.

Who this is for

For: operators whose impressions hold while clicks fall, and founders deciding where the next content dollar goes.

Useful when: you need a system for answer-engine visibility, not another list of tactics.

What must be true

The thesis is five claims, in order. Each one gates the next.

| Stage | Claim | | --- | --- | | Behavior | Queries end in answers, not clicks | | Visibility | Citation replaces position | | Asset | Entities outlast keywords | | Moat | Content must survive summarization | | Channel | The SERP is one surface among several |

01 — The click was the product. The answer is now.

The old contract was simple: rank on page one, collect the click, convert the visit. That contract was already eroding before generative answers shipped — independent click studies had the majority of searches ending without any click at all. AI Overviews, ChatGPT, Perplexity, and Claude finished the job. The query now terminates inside the answer.

This changes the funnel at the root. The legacy chain was rank, click, visit, convert. The new chain is ingest, cite, recall, convert. Your brand either exists inside the answer — named, linked, or paraphrased — or it does not exist at that moment of intent. Traffic still arrives, but it arrives later and warmer: the user who clicks through an AI citation has already been pitched by the machine.

Two consequences follow. First, raw session counts become a misleading KPI; fewer, higher-intent visits is the expected shape, not a failure state. Second, the unit of optimization shifts from the page to the claim. Engines do not cite websites. They cite assertions they can attribute with confidence.

02 — The Citation Stack

Internally we run answer-engine work against a five-layer model called the Citation Stack. Each layer gates the one above it. Diagnose bottom-up; fix the lowest broken layer first.

| Layer | Question it answers | Typical failure | | --- | --- | --- | | 1. Access | Can crawlers and agents fetch it? | JS-only rendering, blocked bots | | 2. Structure | Can a machine parse it without guessing? | Prose walls, no schema | | 3. Entity | Does the model know who is speaking? | Inconsistent naming, no corroboration | | 4. Evidence | Is anything here worth quoting? | Commodity content | | 5. Distribution | Does the claim live where models look? | Single-surface publishing |

Most teams obsess over layer 4 while failing layer 1. An insight that renders only after client-side JavaScript, behind a consent wall, on a slow origin, is invisible to most retrieval pipelines. Audit access before writing a single new word: confirm your robots policy treats GPTBot, ClaudeBot, and PerplexityBot deliberately rather than by accident, and verify your core pages return full content without script execution.

03 — llms.txt, honestly

llms.txt is a markdown manifest at the site root: a curated index pointing agents to your most important content in clean, parseable form. The honest read in mid-2026: adoption by sites is real, commitment by model providers is uneven. No major lab guarantees it is honored.

Ship it anyway. The cost is an afternoon, the downside is zero, and it forces a useful exercise — deciding which twenty pages actually define your company. Pair it with markdown mirrors of cornerstone pages so an agent can pull the substance without fighting your layout. But treat it as hygiene, not strategy. llms.txt is a doormat, not a moat.

04 — Entity authority

Keywords are how pages compete. Entities are how brands compound. Answer engines resolve claims to entities — a company, a person, a product — and weigh whether that entity is a credible source for that claim. Entity authority is buildable, and it has four requirements:

1. One canonical name. Same spelling, same description, everywhere. Variance reads as noise.
2. Machine-confirmable identity. Organization and Person schema in JSON-LD, with sameAs links tying your domain to LinkedIn, Wikidata, Crunchbase, GitHub. Structured data here is not a rich-snippet trick; it is how a model gains confidence about who said what.
3. A claimable territory. A topic narrow enough that your entity becomes the obvious citation. Broad authority is a decade-long project; narrow authority is a quarter-long one.
4. Third-party corroboration. Models weigh consensus. Your claim repeated by people who are not you — podcasts, analyst notes, community threads — is worth more than your claim repeated by you.

At Exiid, every venture leaving Validation gets an entity file: canonical naming, schema baseline, territory definition, corroboration targets. It costs days and pays for years.

05 — The Summarization Test

The hardest question in content strategy now: what survives compression? When a model can reduce your article to two sentences with no loss, the article was a donation to the training corpus. We score every planned asset against five criteria before it gets budget:

1. Proprietary data. Does it contain numbers that exist nowhere else?
2. Named structures. Frameworks with names get cited as-is; generic advice gets absorbed anonymously. A name is an attribution handle.
3. Compression loss. Does a summary destroy the value, or deliver it? Tools, calculators, benchmarks, and templates fail to compress. Definitions and listicles compress perfectly — which is why that layer of the internet is already gone.
4. A falsifiable position. Engines hedge; they synthesize consensus. A sharp, arguable claim is exactly what a hedging machine must attribute rather than absorb.
5. Brand transport. If the claim travels, does the name travel with it?

Three or more passes and the asset ships. Fewer, and we are writing for a machine that will quote us without naming us.

Write the thing the model has to cite because it cannot say it alone.

06 — Distribution beyond the SERP

Answer engines are trained and grounded on more than web pages: YouTube transcripts, podcast feeds, Reddit and niche communities, GitHub, review platforms, newsletters that get quoted and forwarded. A claim that exists on one surface is an anecdote. The same named claim across five surfaces is consensus — and consensus is what engines repeat.

We call the operating motion citation gravity: publish the canonical version on owned property, then deliberately seed the named framework into transcribed and community surfaces. Owned distribution — email, community — is the hedge. It is the one channel where no intermediary model decides whether you are mentioned.

07 — What we measure now

Old dashboards measure a funnel that no longer exists. The replacement set:

• AI referral sessions from assistant domains — small in volume, disproportionate in conversion rate.
• Citation share: a monthly sampled panel of buying-intent prompts run across major engines, scored for whether the venture is named, linked, or absent.
• Branded search volume as the lagging indicator of answer-layer presence.
• Conversion per visit, which should rise as low-intent clicks evaporate. Flat conversion with falling traffic means a real problem; rising conversion with falling traffic means the answer layer is pre-qualifying for you.

Inside the Growth engine this runs as a quarterly loop: sample the prompt panel, score against the Citation Stack, fix the lowest failing layer, reseed distribution. SEO did not die. It moved up a layer — from ranking pages to becoming the source machines trust. The operators who treat that as a system, not a tactic list, inherit the channel.

Read next

• Recon before roadmap — why evidence gates the build plan, the same way access gates citation.
• Is this model worth transferring? — scoring a thesis before committing capital, the discipline this note applies to content.