Generative Engine Optimization

There is no ranking algorithm inside an answer engine in the sense SEOs are used to. A question arrives, the system rewrites it into several sub-questions, retrieves passages against each one, and a language model writes a single answer from what came back. Your brand appears if one of your passages survives that pipeline and the model finds it worth quoting.

That changes what the work is. Position 4 on a results page still got traffic. Passage 4 in a retrieval set that never makes the answer gets nothing at all. The unit of visibility is no longer the page. It is the passage.

So we work at passage level. Every claim that matters to a buying decision gets written as a self-contained statement that survives being lifted out of its page and dropped into an answer with no surrounding context. If a sentence only makes sense after reading the two paragraphs above it, a retriever will either skip it or quote it wrong.

Our own reporting on this is published rather than asserted. Google names and defines query fan-out in its documentation but has never disclosed how the synthesis layer chooses among the pages fan-out retrieves. We write about that disclosure gap directly, because pretending the mechanism is fully known is how agencies end up selling tactics with nothing behind them.

Five independent studies published across 2025 and 2026 landed on the same finding from different directions. Ahrefs, Semrush, SparkToro, Omniscient and Profound all found that off-site brand mentions correlate with AI visibility more strongly than backlinks do.

This is a real break from classical SEO practice. A link is a machine-readable vote. A mention is just your name appearing in text a model was trained on or can retrieve. For a decade the industry treated the unlinked mention as a near-miss worth chasing into a link. In answer engines it is the thing itself.

The practical consequence: a citation-building programme looks less like link outreach and more like getting quoted. Industry roundups, podcast transcripts, comparison articles on sites that rank, directory entries with real editorial text, forum answers where a practitioner names you. None of those need a followed link to move AI visibility.

We build that corroboration deliberately rather than hoping it accumulates. The target is simple to state and slow to earn: when a model is asked who does this work in Toronto, your name should appear in more than one source it already trusts.

The most common GEO pitch in the market is that adding JSON-LD wins AI citations. The evidence does not support it, and we would rather say so than sell it.

Ahrefs ran a controlled test and found a decline in AI Overview citations after schema was added. A separate controlled study on enterprise retrieval found JSON-LD barely moved RAG accuracy, while legible entity pages did. Both results point the same way: structured data helps machines resolve what a thing is, and does very little to make a passage quotable.

We still do a great deal of schema work, and it sits under entity architecture where it belongs. The job there is disambiguation, not citation. Those are different problems with different fixes, and conflating them is how a budget gets spent on markup when the actual blocker was that no page answered the question directly.

If someone quotes you a GEO proposal whose central mechanism is schema, ask what evidence they have. The honest answer is that the published tests point the other way.

Otterly found roughly 73% of sites were blocked from AI citation entirely by crawler-access barriers: robots.txt rules that exclude AI user agents, and content that only exists after JavaScript runs.

That number is worth sitting with. Three quarters of the market has a visibility problem that no amount of content strategy will fix, because the content is never retrieved in the first place. It is also the cheapest problem on this list to solve.

Blocking Google-Extended is the most expensive version of this mistake. It removes you from Gemini and AI Overviews grounding while buying exactly nothing in classical ranking protection, and it is usually switched on by someone acting on a headline about AI scraping.

We check this first on every engagement, before anyone writes a word of content. It takes an afternoon and it decides whether the rest of the programme can work at all.

The first four weeks are measurement. We build a prompt panel from the questions your buyers really ask, run it across ChatGPT, Claude, Gemini, Perplexity and Google's AI surfaces, and record whether you are named, how you are described, and who gets named instead. That is your baseline, and it is the thing every later claim gets measured against.

Then the fixes, in a deliberate order. Crawler access and rendering first, because they gate everything. Entity resolution second, so the engines can tell which company you are. Passage-level content third, written against the specific questions where the panel shows you absent. Corroboration last and continuously, because it is the slowest signal to move.

Reporting runs weekly and shows two things side by side: Search Console impressions, positions and clicks, and the prompt panel. Keeping them separate matters. A gain in classical rankings is not a gain in AI visibility, and reporting that blurs the two is how agencies claim credit for movement they did not cause.

Retainers run from $299 to $1,999 CAD a month, with custom scoping above that. Single audits and scoped projects are available and common.

Crawler-access fixes show up in days. An engine that could not reach you last week can reach you this week, and the prompt panel picks it up on the next run.

Entity resolution takes weeks to months. You are waiting on third-party sources to be recrawled and on the engines to reconcile what they already believe about you with what your site now says. There is no way to force that faster.

Corroboration is the slow one, measured in quarters. Getting mentioned across the sources a model already trusts is a publishing and relationships problem, and it does not respond to spend the way paid media does.

Anyone promising AI visibility in 30 days is describing the first category and quietly implying the other two. We would rather set the expectation correctly at the proposal stage than manage a disappointment in month four.

Classical search has not gone away, and treating GEO as its replacement is a positioning error that costs clients money. Google Search revenue reached $63.3B in Q2 2026, up 17% year over year. The traffic is still there. What changed is how much of it converts into a visit.

Pew logged clicks on results pages with an AI summary at 8%, against 15% without one. Ahrefs measured a 58% decline in clicks on affected queries. The click is getting rarer, not the search.

So the two disciplines run together. Semantic authority and technical SEO earn the classical positions that still bring visits and still feed the retrieval sets AI answers are built from. GEO decides what happens to you once the answer is being written. Buying one without the other leaves a gap that shows up in the reporting within a quarter.

Keyword research and prompt research are not the same exercise, and treating them as one produces a panel full of questions nobody asks a chatbot. People type three words into Google. They write a sentence and a half to an assistant, then follow up twice.

We build the panel from real sources: the questions your sales team gets on first calls, the phrasing in your support tickets, the People Also Ask sets on your commercial queries, and the follow-ups an engine itself suggests when you ask the opening question. That last one matters more than it sounds, because the follow-up is where a recommendation usually gets made.

The panel is deliberately small and stable. Forty to sixty prompts that map to real buying decisions, held constant so movement means something. A panel that grows every month cannot show a trend, which is convenient for an agency and useless for a client.

Access, and a few hours of the right person's time. Search Console and analytics, the CMS or a developer who can merge a pull request, and whoever actually knows the business well enough to say which claims are true.

That last one is the real constraint. The work depends on specific, checkable statements: what you do, who for, where, since when, with what result. Most thin service pages are thin because nobody was willing to commit to a specific claim in writing. We can write the page, but we cannot invent the facts, and we will not publish numbers you cannot stand behind.

Where a claim cannot be verified, it does not ship. That rule costs us a sentence here and there and it is the reason the case studies on this site name only the metrics a client's own Search Console can confirm.