Key Takeaways
Competitive GEO compares your content against the pages AI systems already cite. It separates two problems that often get mixed together: whether your page entered the candidate source pool, and whether it won the citation once it was there.- The paper's strongest practical signals are topic match, candidate position, price information, freshness, and source usefulness.
- Traditional SEO still matters. Google says AI Overviews and AI Mode supporting links require pages that can be indexed and are eligible for snippets; OpenAI separately documents OAI-SearchBot for ChatGPT search result display.
- Formatting alone is a weak first fix. Tables, FAQs, and headings help when they carry facts, criteria, evidence, and answer-ready conclusions.
- The best audit deliverable is not a vague "optimize for AI" memo. It is a prompt set, a citation log, a competitor gap table, and a retest plan.
What The Paper Actually Tested
The What Gets Cited paper studies a narrow but valuable question: when two candidate documents are put into a controlled answer-generation setup, which one is cited first? The authors used a two-document RAG environment, changed one factor at a time, and ran 252,000 trials across six models. That design matters. The paper is not trying to recreate every messy part of a live search product. It removes many outside variables so the citation-selection step becomes easier to inspect. The tested factors include topic relevance, query term coverage, price, specifications, evidence, timestamps, brand signals, position bias, and formatting. The headline result is useful because it is operational. Across models, the strongest and most stable factors are not decorative page patterns. They are whether the candidate source directly matches the question, whether it contains facts the answer needs, whether it is fresh, and whether it appears earlier in the candidate list.| Paper finding | What it means for content teams | Audit question |
|---|---|---|
| On-topic beats off-topic | A page must answer the prompt, not merely the broad subject | Does the section answer the user's exact question? |
| Price beats no price | Commercial pages need comparable facts | Does the page give price, range, package logic, or a clear reason pricing is custom? |
| Recent timestamp beats old timestamp | Fast-moving topics need visible updates | Is the update date visible, and did the facts actually change? |
| Candidate position is powerful | Ranking and retrieval still affect AI citation opportunities | Did the URL enter the source pool before we judged the copy? |
| Formatting-only factors are weaker | Structure should carry evidence, not replace it | Did this edit add facts, proof, comparisons, or just polish? |
Competitive GEO vs Traditional GEO
Traditional GEO asks how to improve visibility inside generated answers. The original GEO: Generative Engine Optimization paper formalized that idea with benchmark tasks, visibility metrics, and content adaptations for generative engines. Competitive GEO asks a sharper question: when your page and a competitor's page are both plausible sources, why does the model cite one and not the other? That is closer to the daily problem for B2B teams. You are not optimizing in an empty room. You compete with SERP leaders, third-party reviews, competitor documentation, community discussions, media pages, and comparison hubs for a limited number of answer citations.| Layer | Traditional SEO question | GEO question | Competitive GEO question |
|---|---|---|---|
| Discovery | Can my page be crawled and ranked? | Can my page enter an AI-usable source pool? | Does my page or the competitor enter the pool more often? |
| Answer use | Does the search snippet earn a click? | Does the answer mention or cite me? | Does the answer cite the competitor before me? |
| Diagnosis | Title, links, technical SEO, content coverage | Entities, answer blocks, evidence, extractability | Competitor facts, first-citation reason, source quality, retrieval position |
| Fix | Improve ranking and clicks | Improve answer readiness | Improve citation win rate against known cited sources |
The Four Citation Gates
The paper's strongest signals map well to four audit gates. They do not explain every answer-engine decision, but they are the right first places to look.1. Topic Match
AI answer engines usually generate from retrieved candidate sources. If the page is generally related but does not answer the prompt, it is a weak citation candidate. A page about "GEO strategy" may be relevant to "best GEO tools for B2B SaaS," but it is not enough if it lacks tools, pricing, criteria, limitations, or selection advice. Google's AI features guidance also keeps this grounded in search: eligible AI features rely on Search systems, supporting links, and content that can be used as part of the search experience. The page must first be understandable as an answer to a query.2. Comparable Commercial Facts
Price showed up as a strong factor in the controlled tests. That makes intuitive sense. If an answer needs to compare products, recommend tools, or explain tradeoffs, the model needs facts it can reuse. A page that only says "contact sales" gives the system less to work with than a page that explains package boundaries, usage limits, who the plan fits, and what drives custom pricing. This does not mean every SaaS company must publish a full price table. It does mean the page should give enough comparison material for a reader and an answer engine to understand how the offer is evaluated.3. Freshness With Real Updates
Freshness matters most when the subject changes quickly. AI search, crawler behavior, Google AI features, and model access policies are all moving targets. The safe pattern is a visible update date plus evidence that the body was actually refreshed. For example, OpenAI's crawler documentation distinguishes OAI-SearchBot, which is associated with search result display in ChatGPT, from GPTBot, which is associated with model training crawling. A page that discusses AI search visibility without version-aware crawler details can become stale quickly.4. Candidate Position
Candidate position was one of the strongest effects in the paper. That does not mean "rank first and everything is solved," but it does mean SEO is not dead. If a page cannot be crawled, indexed, internally linked, or shown with usable snippets, Competitive GEO fixes will not get a fair test. Google's robots meta tag documentation is especially relevant here because snippet controls affect whether content can be used as a direct input for AI Overviews and AI Mode. A page with avoidable snippet restrictions may lose before the citation contest starts.A Competitive GEO Audit Table
Use a 0 to 3 score for each row: 0 means absent, 1 means present but weak, 2 means usable, and 3 means strong and verifiable. The goal is not fake precision. The goal is agreement on what to fix first.| Check | 0-point pattern | 3-point pattern | Why it affects citation |
|---|---|---|---|
| Prompt match | Page covers a broad topic but not the exact question | The opening and relevant H2 answer the prompt directly | Retrieval and generation both reward semantic match |
| Query terms and entities | Missing key entities, category terms, or synonyms | Naturally covers entities, alternatives, and comparison language | Helps the source match the question and answer frame |
| Price or cost boundary | Only says "contact sales" | Gives price, range, package logic, or custom-pricing drivers | Recommendation answers need comparable facts |
| Specs and limits | No feature boundaries, limitations, or audience fit | Lists capabilities, limits, fit, and non-fit cases | The answer can extract sharper conclusions |
| Competitor comparison | Pretends alternatives do not exist | Gives fair criteria, scenarios, and caveats | Comparison prompts need side-by-side evidence |
| Evidence near claims | Marketing claims without sources | Key facts sit near papers, docs, tests, examples, or methodology | Citations need defensible source material |
| Visible freshness | No date or stale body | Updated date plus current facts | Freshness was a stable factor in the paper |
| Technical access | Blocked, noindexed, JS-only body, or no snippet eligibility | Crawlable, indexable, canonical, text-rendered page | No candidate source means no citation |
| Brand/entity consistency | Names, URLs, and descriptions vary across pages | Product, company, schema, social, and third-party facts align | Reduces entity confusion in AI answers |
| Extractable answer blocks | Long paragraphs with no conclusion | Each major section has a standalone answer, table, or checklist | Easier for AI systems to summarize and cite |
First Separate Retrieval Failure From Citation Failure
Many GEO audits jump too quickly into content rewriting. First decide where the failure happened.- Run a stable set of prompts across the target surfaces, such as ChatGPT, Perplexity, Gemini, and Google AI features.
- Record whether your brand, product, or URL appears.
- Record the cited URLs and competitor URLs.
- Check whether your relevant page can be found, crawled, indexed, and summarized in traditional search.
- If the page is not entering the candidate pool, fix technical SEO, internal linking, canonicalization, indexability, and snippet eligibility.
- If the page is retrievable but loses to a competitor, move into Competitive GEO content analysis.
- Retest the same prompts after the fix, using the same platform and similar timing.
Build A Prompt Set That Reflects Real Demand
Do not test one flattering prompt and call it GEO monitoring. A brand may appear for definition prompts and disappear for purchase prompts. It may be visible in English and absent in localized Chinese prompts. A useful first set often includes 25 to 40 prompts across five intent groups.| Prompt group | Example | What to measure |
|---|---|---|
| Definition | "What is competitive GEO?" | Category understanding and source framing |
| Comparison | "GEO vs AEO" or "Convertos alternatives" | Whether the brand enters the right comparison set |
| Purchase | "best GEO tools for B2B SaaS" | Recommendations, citations, price and limits |
| Problem solving | "why is my brand not cited by ChatGPT?" | Whether the page answers the pain directly |
| Brand | "Convertos.ai AI visibility checker" | Brand accuracy and competitor confusion |
Why AI Cites A Competitor Instead Of You
The boring explanation is often the right one: the competitor page gives the model more usable material. It may answer the prompt faster, provide clearer prices, explain limits, include current evidence, use a better comparison structure, or rank higher in the candidate list.| Gap type | Typical symptom | Fix |
|---|---|---|
| Relevance gap | Page is about the topic but misses the exact prompt | Add a direct answer in the intro and the matching H2 |
| Completeness gap | Missing price, specs, limits, audience, or alternatives | Add comparable facts and explain uncertainty honestly |
| Evidence gap | Claims are unsupported or far from sources | Put official docs, papers, tests, examples, or methodology near the claim |
| Competitive gap | Competitor has comparison pages, third-party reviews, or clearer category pages | Build owned comparison assets and improve external profile consistency |
How To Prioritize Fixes
Prioritize fixes by impact, evidence strength, implementation cost, and retestability. Do not rewrite one paragraph for every prompt failure. Start with pages that affect many valuable prompts: category pages, product pages, comparison pages, pricing pages, docs, case studies, and third-party profile pages.| Priority | When to fix first | Example |
|---|---|---|
| High | Affects purchase or comparison prompts, and the page is already discoverable | Product page lacks supported platforms, price boundaries, and competitor comparison |
| Medium | Page answers the question but has weak evidence or stale facts | GEO guide lacks 2026 AI feature and crawler updates |
| Low | Only affects a niche long-tail prompt or cosmetic formatting | Turning a paragraph into a list without adding facts |
- Does this issue affect multiple high-value prompts?
- Is the page already close to the candidate source pool?
- Will the fix add evidence, facts, criteria, or a clearer answer?
- Can we retest citation rate, mention rate, or answer accuracy after the change?
Measurement Plan
Competitive GEO should be reported as a trend, not as a one-off screenshot. A 30-day first measurement window is reasonable because crawling, indexing, AI answer behavior, and result interfaces can lag behind content updates.| Metric | How to record it | Boundary |
|---|---|---|
| Brand mention rate | Prompts where the brand appears / total prompts | Mention does not equal recommendation |
| Citation rate | Prompts citing owned URLs / total prompts | Platforms show citations differently |
| First-citation rate | Times owned URL is the first cited source | Closest to the paper's competitive setup |
| Answer accuracy | Brand-related answers without factual errors | Requires human review |
| Competitor share | Competitor recommendations or citations by prompt group | Split by intent type |
| Fix response | Before/after movement on the same prompt set | Look for repeated movement, not one sample |