In short: There is no ranking algorithm to reverse engineer here. A recommendation is the output of a sequence of filters, and you are eliminated at the first one you fail.
The question gets rewritten into the system's own search terms, candidates are retrieved, most are discarded to fit a budget, the survivors are read, and a few get named.
Diagnosing which step you fail is worth more than any list of tactics, because the fix at each step is different and doing the wrong one changes nothing.
The Sequence, And Why Order Matters
Most GEO advice is a flat list: add schema, write better answers, get more mentions. Flat lists are useless when the steps are sequential, because effort spent on step four is wasted if you are failing step one.
| Step | What happens | What eliminates you |
|---|---|---|
| 1. Interpretation | The question is rewritten into the system's own search queries | You match the user's phrasing, not the system's |
| 2. Retrieval | Candidate documents are gathered | Not indexed, blocked, or not competitive for the rewritten query |
| 3. Selection | Most candidates are discarded to fit the context budget | Your answer is diffuse, or unreadable out of context |
| 4. Synthesis | Survivors are read and an answer composed | Your claim is vague, undated, or contradicted by better sources |
| 5. Attribution | A few sources are named | You were used but a more citable source was credited |
- 01InterpretationEvery page that could plausibly answer is still in play. Your phrasing is already gone.
- 02RetrievalOnly what an ordinary search returns for the query the system wrote for itself.
- 03SelectionMost of those are dropped to fit the context budget. This is where pages die.
- 04SynthesisThe survivors are read and an answer is composed out of them.
- 05AttributionA few are named. Fewer are linked.
Step One: The Question You Answer Is Not The Question Asked
A buyer asks "what should I use to track invoices for a small agency". The system does not search that string. It decomposes the request and issues its own queries, which might cover invoicing software for agencies, small business accounting tools, and comparisons between two named products it already associates with the category.
Two things follow. First, matching the user's exact phrasing is worth much less than it was, because that phrasing is rarely what gets searched.
Second, and more important: if the system does not already associate your product with the category, the query it generates will not contain your name and no amount of on-page work will place you in the results.
This is the hardest step to diagnose, because failing it looks identical from the outside to failing the ones after it.
If you are absent from a category question but present when asked about yourself by name, you are failing at interpretation, not at content. That is a category association problem, and it is solved by other people describing you in that category, not by editing your own pages.
Step Two: Retrieval Is Ordinary Search
The retrieval step is far less exotic than it sounds. Systems use conventional search infrastructure, which means the requirements are the ones you already know: be crawlable, be indexed, be competitive for the query that actually ran.
The failures here are mundane and common. A page blocked in robots.txt for AI user agents. Content that only exists after javascript runs. A page indexed in Google and absent from Bing, which removes you from Copilot entirely. Nothing sophisticated, and worth checking before anything else, because these are cheap to fix and total in effect.
Step Three: Selection, Where Most Pages Die
Retrieval returns more material than fits. Something gets cut, and this happens before any judgement about how good your answer is.
Selection favours passages that are dense, self-contained and clearly located. A page carrying one strong section that answers the question outright beats a longer, better-researched page where the same answer is distributed across several sections. This is genuinely counterintuitive for anyone trained to write comprehensive content, and it is the single most actionable mechanic in this lesson.
The test is simple. Take any section of your page, read it with nothing above or below it, and ask whether it still answers a question. If it needs the preceding paragraph to make sense, it will not survive extraction.
Step Four: What Wins Among Survivors
Now merit matters. Among passages that made it through, the ones used tend to share four properties. None is a ranking factor and all of them are observable in what actually gets cited.
- Specificity. "Reduces processing time by roughly 30% on invoices under fifty lines" is usable. "Dramatically improves efficiency" answers nothing and can be safely dropped.
- Provenance. A claim with a named source and a date can be justified. An unattributed claim has to be hedged or replaced by one that is attributed.
- Corroboration. When independent sources agree, confidence rises. When your page is the only place a claim appears, it is either original research, which is the strongest position available, or unsupported, which is the weakest. The difference is whether you show your method.
- Recency, conditionally. Freshness matters intensely for questions about current state and barely at all for stable definitions. A visible date lets the system judge which case applies. No date forces it to guess, and guessing goes against you.
Step Five: Used Versus Credited
The last filter is the one that shows up in your reporting as a mystery: an answer that clearly draws on your material and names someone else.
Systems tend to attribute to sources that are easy to attribute to. A precise claim you can point at, on a page whose author and date are visible, is easier to credit than the same information dispersed through prose. Where a secondary source has summarised your work more citably than you presented it, the secondary source is what gets named.
The remedy is not to write more. It is to make the specific claims you want credited into things that can be pointed at: stated once, clearly, with a source and a date attached, at a stable address.
On schema markup: Structured data helps by making entities and relationships explicit. It is not a shortcut past any of the five steps above, and no major AI system has confirmed that schema is a direct input to answer selection.
Mark up your pages because it clarifies what you are and improves how you appear in conventional search. Do not expect it to rescue a page failing at selection.
Diagnose Your Own Failure Step
The test plan builder will write the prompts for this from your category and markets. You run them yourself, which is the point.
This is the most valuable exercise in this module. It takes about an hour and replaces guessing with a specific answer:
- Pick one buyer question you should be winning and are not.
- Ask an assistant to answer it, then ask what searches it ran. If your category words are there and your name is not, note it: that is step one.
- Run those exact searches yourself in Google and Bing. If your page is absent, you failed at step two, and the fix is indexing and crawlability.
- If your page appears in ordinary search but never in the answer, paste it into the assistant and ask which passage answers the question. No clear passage means step three, and the fix is restructuring, not more writing.
- If it identifies a passage but competitors still get used, compare their passage to yours for specificity, source and date. That is step four.
- If the answer contains your information under someone else's name, that is step five, and the fix is making your claim more directly citable.
Write down which step failed before you change anything. Teams that skip this reliably spend a quarter improving content while failing at retrieval.
Key Takeaways
- Recommendation is a sequence of filters, not a ranking. You are eliminated at the first step you fail.
- The system searches its own rewritten queries, so absence from category questions is usually a category association problem, not a content problem.
- Retrieval failures are mundane: blocked crawlers, javascript-only content, missing Bing indexing.
- Selection discards material before merit is judged. Dense, self-contained passages survive; comprehensive but diffuse pages do not.
- Among survivors, specificity, provenance, corroboration and a visible date decide what gets used.
- Diagnose the failing step before choosing a fix. The remedy at each step is different and non-transferable.
Check yourself
Before you move on
Not scored, not recorded, and not part of the certificate. Both answers are settled by a sentence in this lesson, and the reasoning appears whichever option you pick.
- 01
At which step of the sequence do most candidate pages get discarded?
- 02
Why does matching the user's exact phrasing matter less than it used to?