Manifesto

AI visibility isn't a content problem.

Why the work that compounds looks structurally different from almost everything being sold under the AEO label — and what we believe about how brands actually become citable.

Most enterprise AEO programmes plateau by month three. Not because the work is wrong, but because the work is content marketing wearing an engineering badge — and content marketing, frankly, is not what AI visibility actually requires.

The industry hasn't admitted this yet. Every agency that sold SEO services for a decade has pivoted into AEO with roughly the same playbook: more articles, more thought leadership, more publishing cadence, dressed up in language about “generative engine optimisation” and “answer engine readiness.” It is the SEO playbook with the LLM acronyms swapped in. It doesn't work. It can't work. And the brands paying for it are about to find out — quietly, at month three, when the dashboard stops moving and nobody can quite explain why.

This firm exists because we think the work has to be done differently. We are an infrastructure firm, not a content agency. We come at AI visibility the way we'd come at site speed, observability, or any other systems problem — diagnose first, instrument second, build the load-bearing pieces in the right order, and let the content compound on top of infrastructure that's actually been built. The methodology compounds. The feature-driven version of the work doesn't.

What follows is what we believe, why we believe it, and what we refuse to do.

Sourceworks manifesto — the foundations of AI search visibility

The Diagnosis

What the industry is getting wrong

AI visibility looks like SEO from a distance. It isn't. The mental model that worked for ten years of search optimisation transfers almost nothing to the answer engines, and the industry pretending otherwise is — honestly — the single biggest reason enterprise budgets are being quietly wasted right now.

A few specific things the industry is getting wrong:

The “publish more, win more” assumption.

SEO rewarded volume. AI doesn't. LLMs are explicitly trained to avoid hallucinating about entities they can't verify — which means content from an unverifiable brand gets weighted lower than no content at all from a verifiable one. The agencies producing 20 articles a month for clients with no Wikidata entry, no consistent schema, no resolved Knowledge Panel — they are, frankly, optimising the wrong layer. The void is louder when you fill it with marketing.

The Submission-Scam Tax.

A growing vendor category is selling “direct submission to ChatGPT” or “AI indexing services for Claude and Perplexity.” Money is changing hands. The services are fictional. ChatGPT, Claude, and Perplexity have no publisher portals. Their indexes are populated through training data, real-time web crawling, and licensed content feeds — none of which a vendor can submit to on your behalf. Any service offering “direct AI submission” is selling something that does not exist. We won't name names here. We will say this: if your current vendor is selling that, your budget is being burned on infrastructure that isn't there.

The dashboard mistake.

A team buys Otterly, Peec, or HubSpot AEO, logs into the dashboard, and calls that measurement. It isn't. A dashboard with no defined success criteria, no rolling averages, no documented baseline, no competitive view, and no agreed targets is just data — and data isn't a story a CFO can read. The result, predictably, is a programme that survives until the third quarterly budget review and then doesn't.

The Reddit avoidance.

Enterprise marketing teams have been told for a decade that Reddit isn't safe for the brand. That advice was reasonable in a world where Reddit fed nothing downstream. In a world where Reddit alone accounts for roughly 40% of citations across major AI platforms, it is the most expensive single piece of marketing advice still being followed. The risk of engaging poorly is real. The cost of not engaging at all is, frankly, higher — and rising every quarter.

The “Team” byline.

Half the enterprise content on the internet is published under “Team,” “Editor,” “Marketing,” or no byline at all. From AI's perspective, that content was written by no one — and AI weights it accordingly. The agencies producing it have not yet absorbed that anonymous publishing, a perfectly fine SEO pattern for fifteen years, has become one of the most expensive habits AEO has rendered obsolete.

Each of these is a different version of the same underlying mistake: treating AI visibility as a content problem when it is, structurally, an engineering one. The work that actually moves the needle is mostly invisible — schema graphs, robots.txt directives, sameAs arrays, sitemap submissions, author entity work, indexing-layer plumbing — and the work that's visible (the content, the publishing cadence, the dashboard) is downstream of all of it. Get the engineering wrong and the content layer can't save it. Get the engineering right and the content compounds.

Seven Positions

What we believe

Seven positions, in roughly the order they show up in the methodology. None of them are novel on their own. What's novel is taking all of them seriously at the same time.

AI visibility is an engineering problem.

This is the load-bearing position. Everything else flows from it.

The reason most enterprise AEO underperforms is that the work that compounds is structural, not editorial. Whether your site's crawlers are getting through. Whether your Organization schema agrees with your Article schema agrees with your Person schema. Whether your sameAs array is bidirectional. Whether your Wikidata entry exists at all. Whether the buyer's actual vocabulary has been mapped against your prompt library. Whether you've earned mentions in the third-party surfaces AI is structurally trained to weight. Whether Bing has crawled your latest pillar piece. Whether your Custom GPT, if you have one, solves a real problem or sits in the store as a record that you built something useless.

None of that is content. All of it has to be in place before content compounds. And almost none of it gets done at enterprises that have hired a “content-first” AEO partner — because the content-first vendor's incentives point at the content, the auditable deliverable, the thing the marketing team can show their CMO this quarter. The infrastructure work is unglamorous, slower, and rarely produces a screenshot anyone wants to share. So it doesn't get done. So the programme plateaus. So the budget gets cut. We've seen the pattern enough times to be able to call it before the engagement starts.

We come at this work in the order it actually needs to be done. Foundation first — the prompt library, the entity layer, the crawlability audit. Authority second — the mention graph, the trust chain, the author entity work. Distribution third — the publishing surfaces AI weights, the indexing infrastructure, the agent-readable layer for the clients it applies to. Measurement runs alongside all of it, because every line of execution has to be defensible at the line-item level. The sequence isn't arbitrary. Each layer compounds the next, and skipping a layer to chase a visible deliverable downstream produces exactly the kind of programme finance cuts.

Third-party consensus is the asset. Your website is the support layer.

The brands AI cites are not the brands publishing the most. They are the brands the rest of the internet is already talking about — and this is the position the industry has had the hardest time absorbing.

AI doesn't build its picture of your brand from your website. It builds it from a distributed consensus: Reddit threads, G2 reviews, Wikipedia entries, podcast transcripts, industry trade publications, comparison sites, conference programmes, expert quotes in tier-2 press. The owned domain is one input. In categories where the conversation is active, it isn't the most important one. The agencies still anchoring content strategy on the company blog are optimising a node in the graph — not the graph itself.

This has uncomfortable implications for how most enterprise PR and content teams measure success. Referring domains and DA scores are SEO metrics. AEO runs on mentions — linked, unlinked, in transcripts, in threads, in pieces that wouldn't pass a traditional PR brief. A single substantive Reddit comment in the right subreddit moves AI visibility further than ten no-follow backlinks from forgotten content farms. A founder appearance on a relevant podcast produces a transcript AI keeps pulling from for years. A piece of proprietary research with a quotable statistic becomes the source AI grounds the same answer in, for as long as the data holds. Those compounding assets are the work. The blog is the support layer.

We tell clients this plainly. Most don't enjoy hearing it the first time. Most absorb it by month two. The teams that don't absorb it are the teams whose programmes will not survive next year's budget review — regardless of who's running them, and regardless of how good the content was.

The buyer's first stop is AI, not the website — and the methodology has to follow the buyer.

For a growing subset of buyers, the journey now starts in ChatGPT or Claude. They ask the category question, get a shortlist of three tools or three providers, and pick from it. Your website never enters the journey. The URL is incidental.

That shift produces a new surface — what we call the Agent Layer — and most enterprises haven't built into it at all. AI tool directories that get cited when buyers ask “what's the best tool for X.” Custom GPTs that put a useful version of your methodology inside the buyer's AI workflow. MCP servers that let users query your product from inside Claude or Cursor. Agent-readable documentation that lets AI agents reason about whether your product fits the buyer's need.

Two things we believe about this layer, both contrarian.

One: most Custom GPTs aren't worth building. The pattern we keep seeing is a brand publishing a “Helpful [Brand] Assistant” with no proprietary methodology behind it, no narrow problem it solves better than generic ChatGPT, no real use case. It sits in the store as a permanent record that the brand shipped something useless. We won't build that, even when the client asks for it. A Custom GPT earns its place when there's a proprietary methodology to package as a tool. Otherwise it's a credibility cost.

Two: MCP is the layer most enterprise SaaS hasn't taken seriously yet, and the early-mover advantage is real and time-bound. For products that buyers would genuinely want to query from inside an AI client — CRMs, analytics tools, project management apps, knowledge bases — shipping an MCP server now puts the brand inside the buyer's workflow at the moment that workflow is being defined. By 2027, every category leader will have one. The brands building this infrastructure in 2026 will be the ones the agents already know how to talk to. The window doesn't stay open.

Authors matter as much as brands.

This is the position most enterprise content teams resist hardest, and it's the highest-leverage single fix in the entire Authority pillar.

AI doesn't weight content by quality. It weights it by trust — and trust is a chain of verifiable connections. A real author with credentials, schema linking that author to the organisation that published them, a topical cluster of content reinforcing what they know about. We call this The Trust Chain. Most enterprise content has none of it. The writing is fine. The chain behind it doesn't exist.

The single most expensive habit AEO has rendered obsolete: publishing under “Team,” “Editor,” or no byline at all. We see this in almost every enterprise audit. Every piece of content, regardless of how good it is, starts from zero credibility because there's no author entity for AI to weight it against. The fix is mechanical: real bylines, real Person schema, real sameAs verification to LinkedIn and Wikidata, real credentials displayed, real knowsAbout declarations. Once the author entity exists, every piece of content that author publishes — on the site, on LinkedIn, on third-party publications — strengthens the entity further. Skip it and every byline starts from zero. Forever.

This applies to senior leaders especially. The CEO who has no Wikidata entry, no consistent identity across LinkedIn and Crunchbase, no credentials displayed on their author page — is, from AI's perspective, an unverifiable person publishing through an unverifiable brand. The category leader who fixes this gets compounding returns. The category leader who doesn't, doesn't.

Indexing for Google is not indexing for AI.

For a decade, “we're indexed in Google” meant “we're indexed.” For AEO, that hasn't been true for a while.

Bing's index feeds ChatGPT Search, Copilot, DuckAssist, and partner integrations across most of the AI ecosystem. Microsoft is OpenAI's largest investor, and the integration runs deeper than a brand partnership — ChatGPT's web search infrastructure runs on Bing's index. The field-wide finding has been consistent: more than half of AI answer blocks cite at least one Bing-discoverable URL. Bing has 3% of search market share. For AEO, it punches at roughly 15-20x its weight — and almost no enterprise team has set up Bing Webmaster Tools, let alone the AI Performance dashboard Microsoft shipped in February 2026.

That dashboard, frankly, is the most underused piece of tooling in the entire AEO stack. It is the only first-party AI citation data available anywhere — for any platform. Which specific URLs are being cited. Which queries are grounding AI answers in your content. In which AI surfaces. Changing week over week. The setup takes an afternoon. The data is irreplaceable. The fact that almost no in-house enterprise team has logged into it is the single clearest indicator that the industry hasn't yet caught up to where the work actually lives.

Combine this with the Submission-Scam Tax we named earlier and the picture is uncomfortable: vendors are charging enterprises for fictional “direct AI submission” services while the genuinely free, genuinely useful indexing infrastructure sits unconfigured. We get clients to fix the free thing in week one. The vendor relationship usually ends in week two.

The methodology compounds. Feature-driven work doesn't.

This is the philosophical position underneath everything else, and the one we'd defend hardest.

Feature-driven AEO work — “let's ship a Custom GPT this quarter,” “let's get a piece into Forbes,” “let's optimise these three articles for AI citation” — feels productive in the moment and produces almost nothing durable. Each deliverable is a one-time spike. The next quarter, the team is back where it started, looking for the next feature to ship. The dashboard hasn't moved structurally. The infrastructure hasn't accreted. The work is, weirdly, all output and no compounding.

Methodology work is the opposite shape. A prompt library, once built well, runs for years and gets refreshed quarterly. A Wikidata entry, once populated, accrues authority every time it's referenced. An author entity, once verifiable, makes every future piece that author publishes weight more. A Trust Chain, once connected, multiplies every mention. A Mention Graph audit, once you've done it, points at the next twelve months of high-leverage work without further input. A measurement framework, once agreed in writing with leadership, protects the engagement through every inevitable bad month.

The methodology has nine concepts in it. We use the same nine names with every client because they describe load-bearing pieces of the work — and load-bearing things deserve names. The Brief. The Verification Stack. The Access Layer. The Mention Graph. The Trust Chain. The Citation Surface. The Indexing Layer. The Agent Layer. The Proof Layer. Some clients find the vocabulary precious at first. By month three they're using it themselves — because the vocabulary is how the work compounds across the organisation. Once everyone is talking about the same nine concepts, the next quarter's priorities pick themselves.

The feature-driven version of this work has no such vocabulary. Which is, weirdly, exactly why it doesn't compound.

Measurement is what makes the work defensible.

The Proof Layer is the difference between an AEO programme that survives its third quarterly review and one that gets quietly cut when the budget conversation starts.

AEO is harder to measure than SEO for a structural reason. There's no SERP. There's no public leaderboard. AI responses vary across users, sessions, regions, and platforms — the same prompt asked twice in a row can return two different answers. SISTRIX's 2026 research found 56-74% of sources cited by AI rotate week-over-week. That volatility isn't a bug; it's a property of the medium. Which means a measurement framework that doesn't absorb the noise produces reports that look like failure every other month — and a programme that's actually working can get cancelled in month four because the wrong chart hit the wrong inbox on the wrong morning.

The default response is to buy a tool, look at its dashboard, and call that measurement. It isn't. Measurement is the framework around the data: which prompts are being tracked and why, what “good” looks like for this specific business, what cadence absorbs the noise without losing the signal, how the report reads to a CMO versus a CFO versus a board. Five metrics. Rolling 12-week averages. Targets agreed in writing at 90 days, six months, and twelve months. Tools triangulated, never trusted singly. An attribution model that names its own limits explicitly and under-claims to keep the rest of the number credible.

This is the work that protects everything else. Without it, the programme is invisible to the people writing the budget. With it, the work defends itself — and the conversation about next year's spend stops being a defence and starts being a planning exercise.

Aside

A note on the vocabulary

We use nine named concepts to describe the methodology: The Brief, The Verification Stack, The Access Layer, The Mention Graph, The Trust Chain, The Citation Surface, The Indexing Layer, The Agent Layer, The Proof Layer. Each one describes a load-bearing piece of the work. Each one describes something the industry has, mostly, not bothered to name — and the things you don't name, you don't manage.

The vocabulary isn't decoration. It's how the methodology compounds inside a client organisation. A CMO who knows that the Mention Graph drives third-party consensus, the Trust Chain makes each mention count for more, and the Indexing Layer is what makes published content retrievable — that CMO can prioritise the next quarter's work without us in the room. A measurement framework that uses the same nine names produces reports the board reads in the same vocabulary the execution team uses. The compounding asset isn't just the methodology; it's the shared language for talking about it.

We didn't invent everything we named. We named everything we found load-bearing.

Position

Where we stand

A short list of what we are, what we aren't, and what we refuse to do.

We are an infrastructure firm.

We come at AI visibility as engineers. Structured data, crawl behaviour, citation graphs, prompt coverage, measurement — these are systems problems, not content problems. We instrument first, build the load-bearing layers in the right order, and let content compound on top of infrastructure that's actually been built.

We are not a content agency.

We don't sell article packages, content calendars, or thought leadership programmes. There are firms that do that well and we'll refer clients to them when content production is what they actually need. Most of the time, it isn't the bottleneck.

We won't build a Custom GPT for a client who doesn't have a real use case.

The temptation is real — Custom GPTs are visible, demoable, and easy to put in a deck. They are also useless when there's no proprietary methodology to package as a tool. A useless Custom GPT is a credibility cost we won't ask a client to pay.

We won't ship an MCP server as a press release.

When the buyer would benefit from accessing the product from inside their AI client, MCP is high-leverage work. When the founder wants the announcement but users don't want the integration, it's a maintenance burden with no payoff. We'll say no to the second version, even when it's the more flattering project.

We won't fake a ranking report.

AEO has no SERP. The consultant who manufactures a “ranking position” to satisfy the SEO-trained instinct undermines the rigour of every other metric in the report. The teams that defend their AEO budgets are the teams that have made the mental shift from rank to citation occupancy and share of voice. We help clients make that shift; we don't paper over it.

We won't sell direct AI submission services.

They don't exist. The vendors selling them are charging for infrastructure that isn't there. We'll show clients how to spot the pattern. We'll redirect the budget into the indexing and agent-layer work that actually moves citations.

We won't claim attribution we can't defend.

Multi-touch helps. It doesn't solve. The honest framing is partial-attribution under-claim — directional confidence, named limits, AI-influenced pipeline as the headline metric rather than a single deterministic number. The consultancy that claims definitive AI attribution is selling a model that doesn't exist, and the moment a CFO asks how the model works, every other number in the report loses credibility.

We're small by design.

Senior-led, working with a deliberately small number of enterprise clients at a time. AEO is a young discipline, and most of the meaningful judgement calls are still human-shaped — the prompt library work, the Mention Graph priority calls, the attribution language, the Custom GPT yes/no. We don't scale by adding junior consultants. We scale by getting the methodology right and keeping the engagement count small enough that the work doesn't drift.

Closing

We started this firm because we watched too many enterprise marketing teams pour budget into AI visibility programmes built like content programmes — and watched the same programmes plateau by month three. The work was wrong because the model was wrong. The model was wrong because the industry hadn't yet decided what AI visibility actually was.

We think it's infrastructure. We think the brands that win the next decade of AI-mediated discovery are the brands that treated it that way early — the ones that built the entity layer, opened the access layer, earned into the mention graph, connected the trust chain, widened the citation surface, wired the indexing layer, shipped the agent layer where it earned its place, and proved every piece of it at the board level.

The methodology compounds. The work defends itself. The buyer's AI finds you where you've built the surface for it to find you, and not anywhere else.

Sourceworks manifesto — the evolution of search from keywords to AI-driven answer surfaces

That's what we believe. That's what this firm exists to do.