---
name: ai-search-explained
description: Authoritative explanation of how AI search actually works — the AI answer chain, where it diverges from classic search, the 4 gates of visibility, the 3 forces (freshness, authenticity, semantic richness), the centroid model, and the three eras of AI visibility. Use to explain grounding, evidence selection, passage selection, query fan-out, and citation mechanics. Built and maintained by Momentic.
version: 2.0.0
---

# ai-search-explained

A teaching skill for explaining how AI search and grounding actually work — at a level deep enough to answer technical questions, debug visibility problems, and make defensible recommendations. Use it when you need to explain *why* AI search behaves the way it does, not just *what* to do about it.

## When to use this skill

Trigger on any of these:

- "How does AI search work?" / "How does ChatGPT (Perplexity / Claude / Gemini / Copilot) decide what to cite?"
- "What is grounding?" / "What's the difference between grounding and search?"
- "Why does my page rank in Google but never get cited by AI?"
- "Why is my competitor cited but not me?"
- "What's the difference between SEO and GEO/AEO?"
- "Does freshness matter for AI search?" / "Does schema matter for AI?"
- "Why was my content excluded from an AI answer?"
- Any question about query fan-out, retrieval, evidence selection, citation, or grounding mechanics.

If the user wants you to *audit their site*, pair this skill with `geo-aeo-readiness` or `agent-ready-checklist` — those use the MCP tools to inspect a real domain. This skill is for *explaining the model*; those skills are for *applying it*.

## TL;DR mental model

Two facts to anchor everything else:

1. **The web index is shared substrate for search and AI.** Both query it. What's changing is the *constraint* applied.
2. **Search optimizes for ranking documents. Grounding optimizes for selecting information that supports an answer.** Same index, two purposes pulling it in different directions.

Translated to optimization disciplines:

| Discipline | Optimizes for | The unit of success |
|---|---|---|
| **SEO** | Ranking pages | A page appears in the top results |
| **GEO** (Generative Engine Optimization) | Being selected as evidence | A passage from your page is cited in the AI's answer |
| **AEO** (Answer Engine Optimization) | Being the answer | Your content is the source the AI quotes from |
| **HEO** (Hybrid Engine Optimization) | Both at once, treated as a unified scoreboard | One score across organic + AI surfaces |

These objectives are not just different — they're increasingly *pulling the index in different directions*. A page can rank well in classic search and still never be cited in an AI answer. The reverse is also true.

**The most common misconception:** treating "AI search" as "SEO with extra steps." It isn't. Grounding has its own pipeline that runs *after* the same retrieval-and-ranking work, with different selection criteria.

The deeper framing: **the index itself is the same shared substrate. The *constraint* applied to it is what's changing**. The same index that was tuned for ranking documents is being asked to support evidence selection, multi-step reasoning, and agentic interaction — each with different demands. A page tuned for one constraint can be invisible under another.

## The 9-stage content pipeline (broad view)

For *content*, the simplest pipeline is 9 stages:

```
Crawl → Render → Index → Interpret → Expand/Fan-out → Retrieve → Rerank → Synthesize → Cite
```

Most teams measure stage 9 (citations / rankings). **Failure lives upstream.** If you fail Crawl, none of Render → Cite matters. The work is figuring out *which stage* you're losing at, then fixing upstream first. This is the spine of `geo-aeo-readiness`.

## The AI answer chain (granular view)

For a *query*, the granular pipeline is twelve stages, splitting at a branch point.

### Common stages — both search and grounding go through these

```
U  → User enters a query
1  → Query understanding         (1a Examination,  1b Understanding)
2  → Query transformation        (2a Normalization, 2b Rewriting, 2c Expansion/Fanout)
3  → Multi-path retrieval        (lexical, semantic, vertical, graph, real-time)
4  → Candidate processing        (dedupe, merge results from retrieval paths)
5  → Multi-stage ranking         (progressively scores and refines candidates)
6  → Eligibility & filtering     (safety, policy, quality thresholds)
```

After Stage 6, the pipeline reaches a **branch point**. The same filtered candidate pool is used for two distinct purposes.

### Search-specific stages — the SERP path

```
7A → Result composition          (organize candidates into structured page with verticals/modules)
8A → Snippet generation          (titles, descriptions, previews)
9A → SERP output                 (ranked links and modules for exploration)
```

This is the path most SEO work targets. The deliverable is a **navigation surface** — links the user clicks.

### Grounding-specific stages — the answer path

```
7B → Evidence selection          (subset that collectively provides sufficient, relevant support)
8B → Evidence construction       (extract & structure passages with attribution → grounding context)
9B → Constrained generation      (answer conditioned on selected evidence, not free-form knowledge)
10B → Cross-check & verification (e.g. Microsoft Critique — evaluates draft against evidence; may re-retrieve)
11B → Multi-model judging        (multiple models validate / challenge / improve, marking what they agree and disagree with)
12B → Answer generation & citation (validated response with citations and, where applicable, actions)
```

This is the path that produces ChatGPT / Copilot / Perplexity / Claude with-search answers. The deliverable is a **synthesized answer** with citations — the user often never clicks.

### The critical observation about Stage 1

**Visibility doesn't start at retrieval (Stage 3). It starts at Stage 1 — query understanding.**

Stages 1 and 2 break into five sub-stages that all run *before* a single document is retrieved:

```
1a → Query examination     (parse: tokens, language, structure, basic signals)
1b → Query understanding   (infer: intent, entities, constraints, context)
2a → Query normalization   (canonical forms, units, casing — "this year" → "2026")
2b → Query rewriting       (vague → executable — "cutest cat toys" → "best cute cat toys 2026")
2c → Query expansion / fanout (one query → many parallel variants covering different intents)
```

**The first thing that actually happens is mathematical:** the raw query is converted into an **embedding** — a vector representation in high-dimensional space. Every subsequent stage operates on the math, not the surface text. Two queries with very different wording but the same intent can pull the same candidates because their embeddings are close in vector space.

If your content doesn't match the *expanded* query forms — not the original user phrasing — it never enters the candidate pool.

### Multi-stage ranking is progressive, not single-pass

Stage 5 (Multi-stage ranking) is frequently misunderstood as "one ranker." It isn't. Ranking is **progressive**: a coarse, cheap ranker runs first; survivors get passed to a more expensive, more rigorous ranker; survivors of *that* get passed to the next, and so on. A page can clear coarse ranking and die at stage 2 or 3.

### Retrieval can loop

Stage 3 (Multi-path retrieval) is not necessarily one-shot. If downstream stages — particularly Stage 7B Evidence Selection — find the retrieved pool doesn't *collectively* support the user's question, the system can **discard those candidates and re-retrieve** with revised parameters. A page can be fetched, judged insufficient as evidence, and the system goes back for more. This is also why retrieval is described as *iterative, stateful, and context-dependent*.

### Personalization and memory shift the candidate pool

Modern retrieval is also context-dependent at the **user** level. Personalization (signals about who's asking — language, location, history) and memory (signals about prior interactions in the same session) shift which candidates surface for a given query. Two users running the *same* query at the *same* moment can pull different candidate pools. This makes "I checked my ranking" an even less reliable measure of citation likelihood than it already was.

## Worked example: "What are the cutest cat toys this year?"

Tracking one query through stages 1–2:

**1a — Examination.** The system pulls signals out of the raw surface:
- "what are" → request for recommendations / list-style answer
- "cutest" → subjective superlative
- "cat toys" → product category
- "this year" → freshness / current-year constraint

**1b — Understanding.** Infers intent at the structured level:
- intent = product discovery
- task = recommend
- entity = cat toys
- subjective_preference = cute / adorable / visually appealing
- freshness = current year
- desired_output = ranked suggestions

Two important inferences happen here: (a) "shopping, not defining" — the user wants suggestions, not a definition of cuteness; (b) "cutest" is not a precise retrievable attribute, so it gets mapped to proxies: *adorable, plush, novelty-shaped, giftable, visually appealing*.

**2a — Normalization.**
- "this year" → 2026
- "cat toys" → `pet_toys > cat_toys` (canonical retail category path)
- "cutest" → subjective ranking preference (a ranking *hint*, not a literal product field)

**2b — Rewriting.** Vague human phrasing becomes executable forms:
- "best cute cat toys 2026"
- "top cute cat toys this year"
- "adorable cat toys 2026"
- "best cat toys with cute designs 2026"
- "popular cute cat toys for indoor cats"

**2c — Fanout.** One query becomes parallel paths covering different intents:

| Aesthetic | Popularity | Engagement | Reviews | Freshness |
|---|---|---|---|---|
| cute cat toys 2026 | best cat toys 2026 | interactive cat toys cats love | cat toy reviews 2026 | new cat toys 2026 |
| adorable cat toys | top rated cat toys 2026 | fun cat toys for indoor cats | best reviewed cat toys this year | trending cat toys 2026 |
| plush cat toys for cats | most popular cat toys this year | best engaging cat toys 2026 | top cat toys to buy 2026 | latest cat toys this year |
| novelty cat toys | | | | |

This is just an *illustrative* fanout. In production, fan-out can produce ~15 variants per visible prompt, and downstream retrieval × ranking yields ~20 candidates per variant — roughly **600 candidate items per visible prompt** before evidence selection ever runs.

**Implication for content creators:** if your page is optimized for *one* phrasing — say "best cat toys 2026" — but says nothing matching *adorable*, *plush*, *interactive*, or *latest*, you compete on one of five fanout paths instead of five. Modular content covering multiple intents within a topic outperforms single-phrase optimization. **The unit of work is the fan-out, not the keyword.**

## The 4 gates of visibility

Content has to pass four gates to be cited. Each gate is independent — failing any one removes you from the answer pool entirely.

### Gate 1 — Discoverable (reachable)
*Can the system find your content?*
- Crawlable URL, sitemap present, no `robots.txt` blocks, clean canonical
- IndexNow push (where supported) — accelerates freshness signals
- Failure mode: AI bots are blocked, your URL is unreachable, or your firewall/CDN challenges agent UAs (the silent killer — robots.txt allows the bot, but Cloudflare bot management or a WAF rule returns 403)

### Gate 2 — Parsable (readable by machines)
*Can the system extract structured meaning?*
- Semantic HTML, structured data (JSON-LD), alt text on images, transcripts on video, clean heading hierarchy
- Failure mode: visually impressive page that's a JS-only blob with no semantic structure

### Gate 3 — Groundable (safe to cite)
*Can the system trust this passage in isolation?*
- Self-contained passages — the meaning doesn't require reading three other paragraphs
- Clear provenance — who said it, when, on what authority
- No contradicting claims on the same page — grounding will deprioritize ambiguous sources

### Gate 4 — Trustworthy (corroborated)
*Is this consistent across sources and time?*
- Entity-graph aligned — the entities on your page match the canonical entity graph (Wikipedia / Wikidata / official knowledge graph)
- Named sources — quotes attributed, data cited
- Multi-surface presence — the same claim on multiple authoritative surfaces
- Consistent across time — your story doesn't change every 6 months

## The 3 forces of visibility

Gates are pass/fail. Forces are what makes you win once you pass.

### Freshness — *pulls* content INTO the result set

Freshness is no longer just a ranking boost. It's an **eligibility signal**. If your information is stale, grounding will probably not select it, **no matter how well-written it is**.

Freshness is fast detection of three things:
- **NEW** content — a page that didn't exist yesterday. Gets a brief attention premium; must qualify quickly.
- **UPDATED** content — primary text, images, schema, video, outlinks. Anything a reviewer model would see differently. Minor typos don't count. Working assumption in industry practice: a meaningful update establishes a ~13-week freshness window before the page reverts to "stale" status.
- **DELETED** content — a URL that returned yesterday and 404s today. Must exit the grounding pool cleanly to avoid stale citations.

Crawl efficiency = information gain per fetch. A site where 99% of crawls return unchanged content burns crawler budget and slows the freshness loop on the 1% that's new.

**Where freshness hits the pipeline:**

| Stage | Effect |
|---|---|
| 3 (Multi-path retrieval) | Surfaces up-to-date content for recency-sensitive queries |
| 5 (Multi-stage ranking) | Boosts newly updated, relevant pages |
| 8B (Evidence construction) | Prefers latest facts when conflicts exist between sources |

**Practical levers:**
- **IndexNow** — open protocol (used by Bing, Yandex, Naver, Yep, Seznam.cz, Amazon) for instantly notifying engines of new/updated/deleted URLs. ~50% of clicked, newly-indexed URLs in Bing SERP originate from IndexNow.
- **499 status codes in your access logs** — non-standard Nginx code meaning *"client gave up waiting"*, which is exactly what AI crawlers do when they hit a slow response. A spike in 499s is a leading indicator of AI-citation cliffs. Most analytics tools don't even monitor this — fix the underlying timeouts and watch citations recover.

### Authenticity — *pushes* content INTO the answer

Named sources, specific data, traceable claims — the qualities the verification stage actively looks for.

**Where authenticity hits the pipeline:**

| Stage | Effect |
|---|---|
| 5 (Multi-stage ranking) | Rewards credible, expert, original sources |
| 6 (Eligibility & filtering) | Removes untrustworthy or spammy pages |
| 10B (Cross-check & verification) | Increases confidence in evidence used for grounding |

A page can be retrieved and selected as evidence and still get cut at 10B if the verifier can't corroborate the claims. Modern verifiers are explicit about disagreement: a model can say *"I agree with paragraph 2; I disagree with paragraph 4 because it contradicts the cited source."* A single contradicted claim can disqualify the surrounding passage even when the rest is correct.

### Semantic richness — *makes content matchable and extractable*

Modular structure, clean passages, schema, entities. Content that's quotable in pieces, not just readable as a whole.

**Where semantic richness hits the pipeline:**

| Stage | Effect |
|---|---|
| 1 (Query understanding) | Better match to expanded intents and fanout paths |
| 3 (Multi-path retrieval) | Stronger semantic matches through embeddings |
| 8B (Evidence construction) | Provides clearer, more extractable facts |

This is why "modular, multi-intent, structured" content outperforms long monolithic articles for AI citation. A 3,000-word article with 8 well-defined H2 sections can be cited 8 times across 8 different queries. The same 3,000 words in a wall of prose gets cited zero times because no clean passage extracts cleanly.

> **Freshness pulls content into the results; authenticity and semantic richness push it into the answer.**

## Pages may rank, but passages have to win

The page is the unit of indexing. The **passage** is the unit of citation.

Within a single page, not all passages are equal. The passages that tend to be selected:

- **Heading + intro** — direct answer in the first 1–2 sentences
- **Definition blocks** — short, declarative, self-contained
- **Evidence-dense data passages** — specific numbers, names, dates
- **Specific how-to steps** — numbered, atomic, executable

The passages that tend NOT to be selected:

- Background context paragraphs
- Methodology / "we did this because" prose
- Transitional filler text
- Marketing fluff

What wins a passage, in five attributes:

1. **Passage clarity** — a single, unambiguous claim
2. **Evidence density** — facts, numbers, names per 100 words
3. **Semantic structure** — under a meaningful heading; in a coherent section
4. **Scope** — the passage answers something the query is asking, fully
5. **Corroboration** — the claim aligns with what other sources say

**Lost in the middle.** Even with infinite context windows, models attend most strongly to the *first* and *last* passages of what they read. The middle is where good content goes to be ignored. **Frontload your best content.** Lead with the answer, then back it up. The "long preamble before the payoff" pattern — common in editorial writing — is structurally penalized by AI selection.

## The centroid model — your brand IS a position in vector space

For an AI system, your "brand" is not your name, your logo, or your messaging. It's the **mathematical average of every chunk of content you've published** — your **centroid** in vector space. AI doesn't *see* your brand; it *calculates* it.

Implications:

- **A clear, distinct centroid → less competition during retrieval, fewer collisions, fewer substitutes.** Distinctive brands are mathematically isolated; AI selects them confidently.
- **Cluster collision** is the silent killer of visibility — when a brand's centroid sits inside a tight cluster with competitors who use the same sources, same best-practice templates, same talking points. AI can't tell you apart and substitutes one for another.
- **Centroid drift** — when content velocity adds chunks that don't belong (off-brand campaigns, partnership content, AI-generated filler), the centroid moves toward the noise. The brand the system has *calculated* is no longer the brand the team thinks they have.

**Diagnostic question:** if you embedded every chunk of content on your site and averaged them, is the result the brand you intend? Or is it diluted by quantity, drifted by partnerships, or collapsed against competitors who write the same things you do?

## The three eras of AI visibility

A clean framing of where the optimization frontier moves over time:

| Era | What mattered | Operational implication |
|---|---|---|
| **GPT-3 era** | Tokens / training-data presence | Be in the corpus. Get crawled, get into Common Crawl, get scraped. |
| **GPT-4 era** | Embeddings / similarity | Be close in vector space. Match the query embedding's neighborhood. |
| **GPT-5+ era** | Traversal of evidence | Be reachable across a connected data layer. Knowledge-graph density and entity connectedness matter more than any single page. |

Brands stuck on the GPT-3 era ("just be in the index") are bringing a knife to a gunfight against brands operating in the GPT-5+ era ("be navigable across a graph"). The shift is from *being indexed* to *being traversable* — from a page to an entity with relationships.

**Training-data presence is not optional, it's the floor.** Industry tests of made-up brands vs real brands on reranker tasks show ~57-point gaps (84.6 vs 26.9 on a 100-point scale). If a brand isn't represented in training data at all, no amount of clever optimization recovers from that floor. Wikidata, Wikipedia, multi-surface presence, citation in third-party authoritative content — these are how training-data presence is built.

## Common misconceptions to correct

When you're using this skill to answer a user, watch for and gently correct these:

| Misconception | Reality |
|---|---|
| "AI search is just SEO with citations." | Same retrieval substrate; different selection criteria after the branch point. SERP-optimized pages routinely fail evidence selection. |
| "If I rank #1, AI will cite me." | Ranking helps — you need to be in the candidate pool. But evidence selection (7B), construction (8B), and verification (10B) all apply additional filters. Ranking ≠ extraction. |
| "Schema is a Google rich-results thing." | Schema is fundamental for grounding — it's how the system knows *what kind* of entity your page is. Without schema, your page is text. With schema, it's a typed entity that can be reasoned about. |
| "Schema by itself is enough." | Industry research shows schema *alone* doesn't reliably move retrieval for general AI systems. What moves the needle is **entity pages with RDF connections** — a knowledge-graph layer the AI can traverse. Schema is necessary; connectedness is what wins. |
| "Freshness is for news sites." | Freshness is now eligibility for *any* recency-sensitive query — and the system decides what's recency-sensitive, not you. "Best [anything] 2026" is recency-sensitive. So is "latest features of [product]." |
| "I can optimize for one main keyword." | Query fanout means one user query becomes 5+ retrieval variants. Single-phrase optimization competes on 1/5 of the surface area. |
| "AI uses everything on my page." | AI uses *passages*, not pages. A page where key facts are buried in long paragraphs gets cited less than the same facts in clean, self-contained sections. |
| "Blocking AI bots in robots.txt is a neutral choice." | It removes you from eligibility entirely. No retrieval, no ranking, no grounding. The trade-off is access to AI training vs. access to AI citation — they're different. |
| "Citation = visibility." | Wrong. **Visibility starts at Stage 1** (query understanding). By the time a citation is awarded at Stage 12B, your page has already won every previous round. Tracking citations measures the *end* of the funnel; the work that earns citation happens 11 stages earlier. |
| "Bigger context windows fix attention." | Models still attend most to the start and end of what they read. Lost-in-the-middle persists even with multi-million-token windows; KV-cache quantization (e.g. TurboQuant) extends *reasoning depth* without solving the attention bias. |
| "Chunk-level optimization is enough." | A team can optimize chunks while the brand centroid drifts. Chunks win retrieval rounds; the centroid governs whether the brand is *substitutable* with competitors. Hold both at once. |

## Controls — what you can choose to expose

AI systems make two distinct decisions about your content:

1. **Access** — *can content be retrieved?* Controlled by `robots.txt`, `noindex`, enterprise permissions, password, paywall.
2. **Use** — *can content be used in snippets/answers?* Controlled by `data-nosnippet`, snippet limits, schema-scoped boundaries.

**No access = no visibility.** If the system can't access the content, it can't determine whether to use it. Access is the gate before use; controls that block access end the conversation before it starts.

### Blunt vs. precision controls

- **Blunt** — `robots.txt`, `noindex`, blocking crawlers. Removes the entire page from eligibility. **No visibility, ranking, or grounding opportunities.**
- **Precision** — `data-nosnippet`. HTML attribute that lets the page stay in the index and rank, but excludes the wrapped section from snippets and AI summaries.

### `data-nosnippet` use cases

```html
<p data-nosnippet>This passage is indexed and helps the page rank, but won't appear in AI summaries or search snippets.</p>
```

| Use case | What it prevents | Why it matters for GEO |
|---|---|---|
| Hide the answer to trigger clicks | Specific portion surfacing | Lets you experiment with which content drives clicks vs. which gets summarized |
| Protect context-dependent content | Nuanced instructions, expert-only guidance | Prevents misinterpretation by LLMs of content that needs surrounding context |
| Safe A/B testing | Experimental copy, variant messaging | Test content without exposing variants in snippets/AI |
| Shield proprietary/premium info | Paid or gated material | Keeps page ranking while blocking evidence extraction of paid passages |
| Control factual grounding | Specs, KPIs, internal metrics | Decide which facts AI can turn into evidence |
| Block tables/lists/specs | High-value extractable structures | Prevents automatic grounding from structured blocks |
| Remove noisy/ambiguous blocks | Fluff, marketing hype, unclear text | Improves semantic clarity and reduces grounding errors |

### Visibility-preserving choice

| Directive | What it does | Impact |
|---|---|---|
| `noindex` | Removes page from search index | **No visibility** — cannot be used in grounding |
| `data-nosnippet` | Blocks specific HTML sections from snippets/grounding | Preserves ranking; hides only selected fragments |

When in doubt, prefer precision over blunt controls. Once you `noindex` a page, you can't be cited from it.

## Practical implications for content creators

If you've read everything above, the actionable shifts follow naturally:

1. **Write for fanout.** Cover multiple intent angles within a single piece, with clear H2s for each. One page wins multiple retrieval rounds instead of one.

2. **Frontload the answer.** Lead with the direct claim, then the proof. Lost-in-the-middle is real even at infinite context.

3. **Make passages self-contained.** Every key claim should be intelligible without reading the previous paragraph. Add the entity, the qualifier, the year — even if it feels redundant for human readers.

4. **Cite specifically.** "Studies show X" is unverifiable; "[Source name] (year, link) reports X" is groundable evidence.

5. **Treat freshness as governance, not a campaign.** Schedule content reviews. Update dates only when content meaningfully changed. Use IndexNow to push updates instantly. 404 retired URLs cleanly. Audit your access logs for 499 status codes — they're the leading indicator that AI crawlers are giving up on slow pages.

6. **Implement entity-grade schema, then go further.** `Organization` and `WebSite` sitewide; `Article`, `Person`, `Product`, `FAQPage`, `HowTo` per page intent. Then use `@id` and `sameAs` to wire your entities into the broader knowledge graph (Wikidata, Wikipedia, official org sites). Connectedness > schema alone.

7. **Choose precision controls.** When you need to limit AI use of specific content, reach for `data-nosnippet` before `noindex`.

8. **Audit on the right axes.** "Where do I rank?" is half the question. "Am I cited? In which grounding queries? With what citation share? Is my centroid distinct or colliding?" is the other half.

9. **Watch the centroid, not just the chunks.** Periodically embed every chunk of your site and look at the centroid. If it's drifted (content velocity from off-brand work, AI-generated filler, partnership pages), the brand AI sees is not the brand you mean to project.

## Closing frame

> *In the AI web, visibility is earned by how well machines can trust, understand, and reuse your content — not just rank it.*

Use this skill to teach the model. Use the companion skills to apply it:

- **Audit a domain or page against this framework:** [`geo-aeo-readiness`](https://momenticmarketing.com/.well-known/agent-skills/geo-aeo-readiness/SKILL.md)
- **Walk the agent-discovery infrastructure:** [`agent-ready-checklist`](https://momenticmarketing.com/.well-known/agent-skills/agent-ready-checklist/SKILL.md)
- **Get specific schema fixes:** [`schema-recommender`](https://momenticmarketing.com/.well-known/agent-skills/schema-recommender/SKILL.md)
- **Audit AI bot access (robots + firewall + CDN):** [`robots-ai-audit`](https://momenticmarketing.com/.well-known/agent-skills/robots-ai-audit/SKILL.md)
- **Score on-page SEO:** [`page-seo-score`](https://momenticmarketing.com/.well-known/agent-skills/page-seo-score/SKILL.md)

These skills call the public Momentic MCP server at `https://momenticmarketing.com/mcp` (tools: `analyze_page`, `analyze_site`).

---

*This skill is built and maintained by [Momentic](https://momenticmarketing.com), an SEO/GEO/AEO/AXO agency. The frameworks here are Momentic's synthesis of public Microsoft / Bing platform research, industry retrieval-systems literature, and ongoing field testing across 100s of mid-market and enterprise clients. The MCP server it depends on (`https://momenticmarketing.com/mcp`) is free for anyone to use. If you want help applying this framework to your site, [book a 30-minute strategy call](https://momenticmarketing.com/contact).*

## Notes for the agent

- **Anchor in the diagram.** When answering a specific question, name the stage. "That happens at Stage 7B — Evidence Selection" is more useful than "the AI decides what to cite."
- **Don't conflate gates and forces.** Gates are pass/fail eligibility. Forces are competitive levers once you're eligible. Confusing them produces vague advice.
- **Use the worked example.** The cat-toys query is mnemonic gold for explaining fanout. Use it.
- **Centroid is the operating model, not a metaphor.** When a user asks "why is my brand getting confused with [competitor]?", the answer is usually centroid collision — same sources, same talking points, no distinctive POV. The fix is differentiated content with first-party evidence, not more content.
- **Watch for "ranking ≠ extraction" confusion.** A user who reports "I rank #1 but never get cited" is hitting Stage 7B failure most of the time. Walk them through evidence selection.
- **Pivot to action via the companion skills.** When the user moves from "how does this work?" to "how do I fix mine?", point them at `geo-aeo-readiness` (full audit), `agent-ready-checklist` (infrastructure), or the specific specialty skills (schema, robots, blog Q&A).
- **Stay calibrated.** This framework is anchored in how grounding-driven AI search works in practice; specifics may vary across engines (Google AI Overviews, Anthropic with-search, Perplexity, OpenAI Search, Microsoft Copilot). Don't claim universal mechanics where the dynamic is engine-specific.