Optimizing for AI Search

• Source: Report research library
• Frameworks applied: RAISE framework (Relevance, Access, Information density, Source authority, Engagement feedback), frozen buyer prompt set, cited-source reverse engineering, three-tier evidence grading

Unusually honest evidence hygiene: every vendor statistic is labelled, and the frozen prompt set plus three-run measurement makes a noisy domain testable.

Known weakness, published as found: Steps 4-7 are long prose blocks; convert each to a short imperative rule set and move the crawler-access grep/robots/WAF procedure into its own clearly gated subsection so it can be read alone. Add a fenced validation gate for step 8 matching the checklists other skills use.

1. 1.Copy the SKILL.md text below, or download the raw file.
2. 2.Create a folder named exactly optimizing-for-ai-search in your agent's skills directory.
3. 3.Save the file inside that folder as SKILL.md.
4. 4.Ask the agent one of the trigger requests below.
5. 5.Check the output against what you already know before it leaves your desk.

• We never show up when buyers ask Perplexity or ChatGPT about our category - what do we fix?
• Build us an AEO program and a way to track our share of AI-generated answers
• Is GPTBot blocked on our site, and what else is stopping LLMs from citing us?

• Build a keyword and backlink plan to rank our blog on Google
• Which AI tools in our martech stack should we consolidate at renewal?

---
name: optimizing-for-ai-search
description: >-
  Optimizes site content, crawler access, and third-party entity footprint so answer engines
  cite the brand, using the RAISE framework, and sets up honest measurement of share of
  AI-generated answers. Use when the user says AEO, GEO, answer engine optimization, LLM SEO,
  "get cited by ChatGPT", "we don't show up in Perplexity", "AI Overviews are eating our
  traffic", GPTBot blocked, "how do we rank in AI search", LLM citation tracking, or share of
  AI answers. Use this skill whenever the task involves being visible inside AI-generated
  answers rather than in classic search result rankings, even if the user calls it SEO. Do NOT
  use for keyword-and-backlink technical SEO programs, for AI-generated outbound email (see
  deploying-ai-sdr-programs), for auditing purchased AI tooling (see auditing-gtm-ai-stack), or
  for internal agent workflow design (see designing-agentic-revenue-workflows).
metadata:
  version: "1.0"
---

# Optimizing for AI search

One job: make a brand's entity and content retrievable and quotable by answer engines, then measure citation share against a fixed prompt set. Paid search, classic ranking work, and content production at volume are out of scope.

## Set expectations before any work starts

The mechanics of answer-engine optimization are credible and low-risk. The measurement in circulation is not. Every quantified AEO claim available traces to unaudited agency client anecdotes, so build the practice and discount the numbers.

Specifically, refuse to forecast on these two figures and say why: **73% of B2B buyers use AI tools in purchase research** has a named source but no published sample or method (Loganix via PR Newswire, cited by [Deloitte Digital](https://www.deloittedigital.com/us/en/insights/perspective/accelerating-b2b-sales-agentic-ai.html)), and the widely repeated claim that answer engines will influence **10–20% of B2B software purchase decisions** has no study, sample, or method behind it ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)). Use them to justify starting the work, never to size a pipeline target.

The one number with a stated sample worth acting on: in an audit of **50 enterprise SaaS websites, 68% were inadvertently blocking at least one major AI crawler** ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)). Date, selection method, and crawler list are not given, so treat it as a prior that access is broken more often than not — which is why access verification comes before content work.

## Workflow

Copy this checklist into your reply and tick items as you complete them:

```
- [ ] 1. Build the buyer prompt set (the unit of measurement)
- [ ] 2. Baseline current citation share against that set
- [ ] 3. Verify crawler access and fix blocks — hard gate
- [ ] 4. Rewrite relevance signals into extractable factual claims
- [ ] 5. Raise information density against the entity checklist
- [ ] 6. Build the third-party source-authority footprint
- [ ] 7. Run engagement feedback and reverse-engineer cited sources
- [ ] 8. Re-measure, fix, re-measure; only report a change after two clean runs
```

Steps 3 through 7 are the RAISE framework — Relevance signals, Access verification, Information density, Source authority, Engagement feedback ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)). Run access before relevance in practice, because rewriting copy that no crawler can reach changes nothing.

**1. Build the buyer prompt set.** Write 20–40 prompts a real buyer would type, covering four types: category definition ("what is FP&A software"), fitness-for-segment ("best financial planning tool for manufacturing companies"), head-to-head ("compare Prophix vs Planful vs Workday Adaptive"), and objection or constraint ("FP&A tools with NetSuite integration and SOC2"). Freeze this set and version it, because answer-engine outputs are non-deterministic and a changing prompt set makes every trend line meaningless.

**2. Baseline citation share.** For each prompt, run it on each target engine and record: whether the brand is mentioned, whether it is cited with a link, the position among named vendors, and the full source list. Run each prompt three times and record the mention rate across runs, not a single result, because the same prompt returns different source sets on repeat calls. Store as one row per prompt-engine-run.

**3. Verify crawler access.** This step is exact and non-negotiable, because it is the only part of AEO with a binary pass condition. Do all four:

- Grep server logs for crawler signatures `GPTBot/1.0`, `ClaudeBot/1.0`, and `PerplexityBot/1.0`. Absence of hits is the finding, not a null result.
- Check `robots.txt` for overly restrictive rules and add explicit `Allow` directives for those agents, because a blanket disallow with no agent-specific allow is the most common block.
- Check WAF and bot-management configuration and whitelist the AI crawlers, because security tooling blocks them without appearing in `robots.txt`.
- Check rendering: minimize JavaScript dependencies for primary factual content, because progressively rendered copy may never enter the crawler's view.

Re-grep logs after the fix and confirm crawler hits appear. Only proceed to content work once at least one hit per target agent is present in the logs, because every downstream step's effect is unmeasurable while access is blocked.

**4. Rewrite relevance signals.** Convert positioning copy into expert answers that state what the product is, who it serves, what it does, and its specific capabilities. Answer engines retrieve solutions, not superlatives.

```
Ineffective: "Our platform empowers finance leaders with next-generation insights."
Effective:   "Prophix is an FP&A platform that automates budget consolidation for
              mid-market companies, with native ERP integrations and AI-powered
              forecasting capabilities."
```

Apply the pattern, do not copy the example. The test for each rewritten sentence: it should survive extraction as a standalone quoted claim with the subject named, because engines quote sentences out of their surrounding page.

**5. Raise information density.** Build a rich, contextually connected set of related terms, concepts, and use cases and embed them throughout the relevant pages — richer, not merely longer. Cover deliberately: category terminology, integrations by name, use cases, named competitors, compliance certifications, security posture, and the metrics the buyer manages. The worked example checklist covers "financial planning & analysis software", "cash flow forecasting and budget variance analysis", "integrations with Sage, Microsoft Dynamics, and NetSuite", "manufacturing financial modeling", "alternatives to Oracle Planning and Anaplan", "SOC2 compliance", and "KPIs like working capital ratio, DSO, and cash conversion cycle" ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)). Build the equivalent list for the brand's own category and audit page coverage against it; this step is judgment work and the list should reflect the actual buying conversation, not a keyword tool's output.

**6. Build source authority.** Third-party validation and digital PR run concurrently with product, blog, and comparison-page work rather than after it. Target five source classes: analyst reports (Gartner Magic Quadrant, Forrester Wave), review platforms (G2, Capterra) with verified reviews, professional forums (Reddit, Quora), trade publications, and partner marketplaces. The claim that brands appearing in at least 5 authority sources see 2.7× higher mention rates is attributed only to "our data" with no sample, date, or method ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)) — use five sources as a planning target and label the multiple as a vendor assertion if the user asks for the expected lift.

**7. Engagement feedback and source reverse-engineering.** Two activities:

- Systematically prompt the answer engines, via employees, customers, and partners, with the frozen prompt set. One agency reports a 30-participant program producing a 4× visibility increase, with no baseline, duration, or query count given ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)) — run the activity because it is cheap, and do not promise the multiple.
- Reverse-engineer cited sources: run a target buyer prompt in Perplexity, read the displayed sources, classify each as analyst report, specialist blog, comparison site, or other, and mirror the source types that actually appear for that prompt. Source-type mix differs by prompt category, so classify per prompt rather than building one global list.

**8. Re-measure and loop.** Re-run the full frozen prompt set on the same cadence (monthly is sufficient; weekly re-runs mostly measure noise). Compute mention rate and cited-with-link rate per engine. If the metric moved less than the run-to-run variance observed in the baseline, the change is not evidence — apply the next fix and re-measure. Only report a directional change after two consecutive runs move the same way, because single-run movement inside a non-deterministic system is indistinguishable from sampling.

## Output format

Use this exact section order, because the citation-share table is the artifact that gets compared across months. Prose is yours.

```markdown
# Answer-engine visibility — <brand>, <period>

## Citation share
| Prompt | Engine | Mentioned (of 3 runs) | Cited with link | Position among vendors | Prior period |
|---|---|---|---|---|---|

## Access status
Per crawler agent: log hits found, blocks identified, fix applied, post-fix hits confirmed.

## Content changes shipped
Page, change type (relevance rewrite / density / new comparison page), date shipped.

## Third-party footprint
Source class, asset, live date, whether it appeared in any engine's source list.

## Cited-source analysis
Per prompt: the source types the engines actually retrieved, and the gap versus our footprint.

## Evidence notes
Every claim labeled Study / Vendor claim / Unsourced assertion. No forecast built on assertions.
```

## Diagnostic questions to force scope

Ask these four before committing to a program, because they separate a measurable AEO engagement from an unbounded content project ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)):

1. Which buyer prompts and category questions matter, named explicitly?
2. Which sources do answer engines actually retrieve and cite for those prompts?
3. What must change on the site and across the wider web?
4. How will visibility connect to engaged evaluation, qualified meetings, and pipeline?

If the user cannot answer 1, do step 1 of the workflow before anything else. If they cannot answer 4, state plainly that no credible attribution model exists for AI-answer influence yet and scope the program to leading indicators — citation share, cited-with-link rate, and third-party footprint coverage — rather than to pipeline.

## Gotchas

- Access is broken far more often than teams assume: 68% of 50 audited enterprise SaaS sites were blocking at least one major AI crawler ([Powered by Search](https://www.poweredbysearch.com/blog/aeo-llm-seo-best-practices/)). Check the WAF as well as `robots.txt`, because bot-management rules block agents invisibly to anyone reading only the robots file.
- A single prompt run is not a measurement. Answer engines return different source sets on repeated identical prompts, so a brand can appear and vanish with no change on the site. Three runs per prompt minimum, and report mention rate.
- Changing the prompt set breaks the trend line permanently. Freeze and version it, and add new prompts as a separate cohort rather than editing existing ones.
- Mention without citation is a different problem from absence. A brand named in prose but not linked has an entity-recognition win and a retrieval-source gap, and the fix is third-party footprint, not on-site copy.
- Unedited marketing prose is actively counterproductive here, because superlative-heavy sentences carry no extractable factual claim and engines quote sentences out of page context.
- The lexical patterns that hurt AI-generated outbound also hurt extraction: "delve / leverage / synergize" vocabulary measured −14% on cold-email reply rate ([Digital Applied](https://www.digitalapplied.com/blog/ai-sdr-real-performance-100k-email-analysis-2026)), and the same abstraction removes the named entities an answer engine needs. Write concrete nouns.
- Every headline AEO percentage in circulation is an agency assertion. Treat crawler access, entity-dense factual copy, and third-party citation footprint as the deliverables, and refuse to attach a pipeline forecast to them.
- Comparison pages naming competitors are the highest-yield asset for head-to-head prompts, because those prompts retrieve comparison-site sources specifically; check the cited-source mix for a head-to-head prompt before deciding what to build.

What does the Optimizing for AI Search skill do?

Makes a brand retrievable and quotable by answer engines using RAISE, then measures citation share honestly

Where does the Optimizing for AI Search skill come from?

Report research library. It was written by The Revenue AI Report against a 12 criterion quality rubric and graded in an independent scoring pass.

Why was the Optimizing for AI Search skill chosen for this library?

Unusually honest evidence hygiene: every vendor statistic is labelled, and the frozen prompt set plus three-run measurement makes a noisy domain testable.

When should the Optimizing for AI Search skill not be used?

Do not use it for: Build a keyword and backlink plan to rank our blog on Google Or: Which AI tools in our martech stack should we consolidate at renewal?

How do I install the Optimizing for AI Search SKILL.md file?

Download the file, create a folder named exactly optimizing-for-ai-search inside your agent's skills directory, and save the file inside it as SKILL.md. The agent loads it when a request matches the description.