AI search optimization: how a site ranks in ChatGPT, Perplexity, and Google AI Overviews

A figure often cited is that AI search today drives less than 1% of traffic for most websites. That is true — direct traffic from ChatGPT, Perplexity, or Gemini is a small part of the picture. But that figure measures only one dimension: visitors who arrive through an AI system. AI visibility today happens elsewhere too — in Google AI Overviews that appear in more than half of all searches and sit above the organic results, in citations a user remembers before ever running a search, in decisions made without a click. The site architecture that ranks in AI systems is not a marketing campaign with a quarterly ROI — it is infrastructure that is built once, used long term, and ranks later more cheaply than it would earlier. The question is not "should the SEO strategy be redirected" — the question is "should AI-readable infrastructure be added to the site now or later, at a higher cost." This guide breaks down AI search optimization through three disciplines — architecture, content, and authority — that form the core of every project in the Praxidea Canon methodology.

Why classic SEO is not enough for AI search

AI search optimization is the discipline that prepares content and infrastructure for citation in generative search systems — ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude. Classic SEO targets the click on the blue organic result; AI optimization targets the appearance of content as a source within an AI answer. The disciplines overlap at the base (the same site that ranks classically usually also becomes a candidate for AI citation), but they differ in mechanics: AI systems do not value the same signal in the same way, nor are all AI systems identical to one another.

The industry formalized the term in the academic paper GEO: Generative Engine Optimization (Aggarwal et al., arXiv:2311.09735, first version November 2023, current version June 2024). The paper defines GEO as "the first paradigm to help content creators improve the visibility of their content in generative engine responses" and documents that systematic optimization increases visibility by up to 40% in AI answers. The Wikipedia GEO entity, last updated May 2, 2026, uses the same formal definition and recognizes the discipline as an active field.

In practice, the terminology is not yet settled. GEO — Generative Engine Optimization — is the most common umbrella term and covers the discipline as a whole. AEO — Answer Engine Optimization — focuses on optimization for systems that generate direct answers (Google AI Overview, featured snippet, FAQ extracts). LLMO — Large Language Model Optimization — appears in some agency contexts as a broader label for AI optimization with an emphasis on language models as the target medium. In this guide GEO is used as the primary term, and AEO as a related subset.

The difference between classic SEO and AI optimization is structurally a difference of goal and signal. Classic SEO optimizes for ranking — position in the list of results. The signals are known and stable: site architecture, content that answers intent, backlinks. AI optimization optimizes for citation — the appearance of content as a reference source in a generated answer. The signals overlap, but the priorities differ: AI systems strongly value the structure of content (how "extractable" the content is as a self-contained block), the entity (whether the brand is clearly recognized as the subject of the topic), and intent coverage (how many sub-queries the content answers). Word count carries less weight than in classic ranking.

A common argument against AI optimization goes like this: AI search today drives less than 1% of total traffic for most websites — it is too early for a strategic shift. The data is accurate as a current snapshot, but the argument does not hold for three reasons. First, Google AI Overviews appear in more than 50% of searches (source: Semrush AI Overviews Study, cited in the Search Engine Land analysis of GEO from February 2026) — which means AI optimization is already happening above the organic results, in Google itself. Second, AI-inspired traffic is structurally invisible in analytics: a user who asks ChatGPT, remembers the brand, then searches directly, shows up in analytics as "direct" or "organic Google" — not as "AI." Third, infrastructure investments (structured data, llms.txt, the entity graph) are low cost, additive, and cumulative. The choice is not "redirect SEO" versus "keep SEO" — the choice is "add AI-readable infrastructure now" versus "add it later at a higher cost."

Architecture: how AI systems read a site

AI search optimization begins with architecture — the infrastructure layer that determines whether an AI system can even read, understand, and cite the content. Architecture comprises four components: structured data (schema markup) that explicitly describes what is on the page; an llms.txt file that gives AI systems structured guidance about the site's content; AI crawler access that governs which bots are permitted to index the content; and an entity graph that links the brand to its public reference points. Without correct architecture, content stays invisible to an AI system even when it is high quality.

Structured data and Google's explicit position

Structured data — JSON-LD schemas of type Article, FAQPage, BreadcrumbList, Organization, Person — gives the search engine and the AI system explicit information about a page's content. Without schemas, the system has to guess from prose. With schemas, the page becomes a candidate for rich results (FAQ extracts, sitelinks, featured answers) and — decisively, in 2026 — a candidate for citation in an AI answer.

In its official documentation (source: Google Search Central, AI features and your website, updated December 10, 2025), Google is clear on one point: there is no additional technical requirement for appearing in Google AI Overviews and AI Mode. Quote: "There are no additional technical requirements. You don't need to create new machine-readable files or markup to appear in these features. There's also no special schema.org structured data you need to add." This means that schema for Google AI Overviews is the same as schema for classic ranking — useful, but not specifically prescribed.

A key distinction: Google's position applies to Google AI Overviews and AI Mode. For other AI systems — ChatGPT, Perplexity, Claude, Gemini — structured data remains mechanically relevant. These systems use structured data for entity detection, content-type classification, and FAQ-pair extraction. A site that emits the full stack (Article + FAQPage + BreadcrumbList + Organization + Person) is in a better position for citation in non-Google AI systems. Schema is not harmful for classic ranking either — it remains the industry standard for a technical foundation.

Per an internal review of the top 10 SR results for AI-related searches (Praxidea audit, April 2026), none of the competing sites emit the full JSON-LD stack; most have not even a basic Article schema. A site that implements schema from day one captures SERP features without direct competition.

The llms.txt protocol

An llms.txt file is a text file in the site's root folder that gives AI systems structured guidance about the site's content. The protocol was proposed in 2024, and since then it has been honored to varying degrees by GPTBot, PerplexityBot, Claude-SearchBot, and some other AI crawlers. llms.txt is not a formal standard like robots.txt — it is an evolving convention. But in 2026, a site without an llms.txt file has failed to use a low-cost mechanism that explicitly tells AI systems how to navigate its content.

The file is placed at /llms.txt (the domain root) and contains structured Markdown with sections describing: the site's identity, its primary purpose, key content nodes (the URLs of the most important pages), methodology documentation, and specific sub-pages the AI system should recognize as reference points. The format is deliberately simple — neither YAML nor XML, but Markdown with # headings that all modern AI systems parse.

A minimal example of the structure of an llms.txt file:

# Site name

A short description of the agency or brand in one to two sentences — who, what, where.

## Basic details
- Name
- Founder
- Headquarters
- Areas of work
- Content language

## Methodology
A short description of the methodological framework and a link to the full documentation.

## Key content nodes
- https://domain.com/page-1 — description
- https://domain.com/page-2 — description

## Content author
A short description of the author and a link to the author page.

Praxidea applies llms.txt on its own site. The file at praxidea.com/llms.txt is publicly available and covers the agency's identity, the Praxidea Canon methodology, primary markets, language, services, excluded industries, pricing policy, AI search as the target medium, key content nodes (all the site's main pages), published articles, and the content author. This structure is not hypothetical — it is applied on a site that simultaneously undergoes classic Google optimization and AI citation.

A complementary convention — ai.txt — is developing in parallel with llms.txt, with a focus on opt-out signals (whether the site permits AI training) rather than reading guidance. ai.txt adoption is lower than llms.txt, but some crawlers already read it. For a site that wants to cover both layers, having both files in the root is a low-cost step.

The AI crawler ecosystem — a three-layer structure across all major providers

The AI crawler ecosystem is not monolithic. Each of the three major providers — OpenAI, Anthropic, Perplexity — operates several crawlers, each with a different role and a different relationship to robots.txt rules. Google adds its own layer via the Google-Extended signal. Understanding this three-layer structure enables precise control: a site can allow AI crawlers that index for search while blocking those that gather data for model training.

OpenAI operates three crawlers, documented on the OpenAI platform docs site:

• GPTBot — gathers content to train OpenAI models. Respects robots.txt.
• OAI-SearchBot — indexes content for the ChatGPT Search feature. Respects robots.txt.
• ChatGPT-User — responds to specific user queries in ChatGPT; does not index preemptively. Respects robots.txt in most contexts, but may access content when a user asks directly.

Anthropic officially documented a three-bot framework in 2026 (source: support.claude.com/en/articles/8896518, also covered in the Search Engine Land analysis from 2026):

• ClaudeBot — gathers content to train Claude models. Respects robots.txt.
• Claude-User — responds to specific user queries within a Claude conversation. Limited robots.txt adherence.
• Claude-SearchBot — indexes content for the Claude search feature. Respects robots.txt.

Perplexity operates a two-part structure (source: docs.perplexity.ai/guides/bots):

• PerplexityBot — indexes content for Perplexity search. Respects robots.txt.
• Perplexity-User — responds to specific user queries and ignores robots.txt ("since the user requested the fetch").

Google-Extended is a signal Google uses for AI training (Bard, Gemini). Google's core crawler (Googlebot) behaves as before; Google-Extended is a separate directive in robots.txt that applies to AI training.

Practical consequence: a site that wants to be found in AI search but not in AI training can allow the search bots (PerplexityBot, OAI-SearchBot, Claude-SearchBot, Googlebot) and block the training bots (GPTBot, ClaudeBot, Google-Extended). The robots.txt rule then distinguishes between user-agent strings:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: Google-Extended
Disallow: /

This is a nuance that most Serbian sites do not implement in 2026 — they treat AI crawlers as a monolithic set. The distinction takes little effort and enables precise control over presence.

The entity graph — sameAs as the bridge to the Knowledge Graph

The entity graph is a network of reference points that connect a brand to its identities in other knowledge systems. In JSON-LD this is implemented through the sameAs field in an Organization or Person schema — an array of URLs pointing to the same entities on other platforms. When an AI system sees that the brand on a site has a sameAs link to a Wikidata entity, a CrunchBase profile, a LinkedIn company page, a Google Business Profile, it builds a confident graph between those reference points. That graph is what Google Knowledge Graph, ChatGPT, and Perplexity read when they generate an entity summary.

The practice of building an entity graph for a Serbian brand:

1. A Wikidata item — the foundational element. Wikidata is a knowledge structure read simultaneously by the Google Knowledge Graph and by all major AI systems. The item contains the brand's identifiers, founders, headquarters, and areas of work.
2. A CrunchBase profile — relevant for agencies and B2B brands. CrunchBase is referenced directly in Google's entity signals.
3. A LinkedIn company page — a standard sameAs node.
4. Professional directories (Clutch, Sortlist, GoodFirms for agencies; industry catalogs for other lines of business).
5. Author profiles — ORCID for researchers and content authors, GitHub for technical firms, LinkedIn for individuals.

All these nodes are connected through the sameAs array in Organization (for the brand) and Person (for named authors). When an AI system performs entity resolution — the process of linking a reference in text to a canonical entity — the sameAs network shortens the path from "a mention in an article" to "the canonical entity" to at most one hop.

Query fan-out decomposition — how Google AI Mode breaks down a query

Google AI Mode and, to a lesser degree, Google AI Overview do not answer a user's query directly. Instead, the system breaks the query into 8 to 12 parallel sub-queries (Google AI Mode typically 9–11), extracts the best answers for each sub-query from different sources, and synthesizes the final answer. Google officially names this process query fan-out (source: Google Search Central, AI features and your website, December 2025).

The consequence for content is direct. A site whose single article answers 8 of the 11 sub-queries generated for a given umbrella search is cited 8 times in the AI answer — not once. A site that answers 3 of 11 is cited 3 times, even though both pages rank classically for the same umbrella search. Query fan-out is, according to the analysis by Aleyda Solis, one of the least understood AI-visibility mechanics; per an internal Praxidea review (April 2026), in the Serbian market almost no one addresses fan-out explicitly.

An example of decomposition for the umbrella query "best SEO strategy for a small business in 2026":

1. What is an SEO strategy?
2. How does SEO differ for a small business compared to enterprise?
3. Which ranking signals matter most in 2026?
4. How long does an SEO strategy take to show results?
5. Should a small business hire an agency or a freelancer?
6. How much does SEO cost for a small business in 2026?
7. Does AI search change SEO strategy for a small business?
8. What tools does an SEO team use in 2026?
9. How do you measure the results of SEO work?
10. Is local SEO worth it for a small business?
11. What mistakes do small businesses most often make with SEO strategy?

An article that would successfully rank for the umbrella query, to be cited as much as possible, should structurally answer 7 to 9 of these sub-queries — not just the umbrella question. Content is written for the fan-out, not for the umbrella query.

Content: how AI systems cite pages

In 2026, content is not valued by length and keywords alone. AI systems extract content in specific formats and prefer structures that are easy to "pull out" as self-contained blocks. Three mechanics determine whether content gets cited: answer capsules (short direct answers, 40–75 words), self-contained paragraphs (which retain their meaning without surrounding context), and E-E-A-T signals calibrated for AI systems (named authorship, primary sources, verifiable expertise).

Answer capsules — a direct answer in the first 75 words

An answer capsule is a block of text of 40 to 75 words that directly answers a specific question, structured so it can be cited on its own. AI systems preferentially extract such blocks — the first paragraph of a section, a section excerpt, or the first sentence under a heading that matches the query. The Search Engine Land analysis from February 2026 identifies self-contained paragraphs as one of the strongest structural factors correlating with citation.

The practice in this guide opens every section with an answer capsule: the first sentence directly answers the section's umbrella question, the second and third provide context, then the in-depth part follows. When an AI system performs fan-out decomposition and needs a quick answer for a single sub-query, it looks first at exactly these opening sentences of sections. A site that opens each section with a generic introduction ("In this part of the article we will look at...") effectively excludes the section from AI citation.

E-E-A-T for AI — named authorship and primary sources

E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) is the framework Google uses to assess the quality of a source. In 2026 the framework also applies to AI citation, but with different weights. AI systems especially value:

• Named authorship — an article must have a visible author with a biography and a verifiable identity. An author without a sameAs link (LinkedIn, ORCID, GitHub) signals to the AI system that the entity "author X" is not connected to the public knowledge graph.
• Primary sources — a citation of a primary document (official Google documentation, an academic paper, a legal text, direct research) is treated by AI systems as a signal that the content has been verified. Secondary sources (other articles on the same topic) carry less weight.
• Citation dates — a source given without a date is treated by an AI system as unverified. A citation in the format "(source: Google Search Central, updated December 10, 2025)" carries more weight than "(source: Google Search Central)" without a date.
• Verifiable first-hand experience — an article that draws on real practice (its own projects, documented cases, measurable data) ranks better than an article that operates only on definitions.

Practical consequence: an article without a named author with a biography, without primary sources, and without experiential references functions in classic SEO as a weak signal — in AI citation it functions as almost no signal at all.

Platform differences in citation — ChatGPT, Perplexity, Google AI Overview, Claude

AI systems differ from one another in how they gather and cite content. Understanding these differences makes it possible to target optimization per platform.

ChatGPT Search — a feature OpenAI launched on October 31, 2024 (covered in Search Engine Land and MIT Technology Review, October 2024). The system uses a fine-tuned GPT-4o model combined with third-party search providers — Microsoft Bing confirmed as one of them. Publishing-house partnerships (Reuters, The Atlantic, Le Monde, Financial Times, Axel Springer, Condé Nast, TIME) supplement open search. Citations are displayed inline and via a Sources button that opens a side panel with references. For ChatGPT, structured content with clear authorship and a publication date is decisive.

Perplexity — uses its own crawler plus third-party search providers. Perplexity displays citations immediately behind each assertion in the generated answer, which makes it the most transparent AI system in terms of attribution. Perplexity's favorite is research-style content that is dense with facts and citations — an article that looks like a summary of multiple sources with cross-citations has a greater chance of being cited. PerplexityBot indexes with robots.txt respect, while Perplexity-User performs user-requested queries and ignores robots.txt (source: docs.perplexity.ai/guides/bots).

Google AI Overview and AI Mode — Google uses its own Search index and the Gemini model to generate answers. The AI Overview appears in more than 50% of searches (Semrush AI Overviews Study, cited in Search Engine Land). Google's official position (Google Search Central, December 2025) is that there are no additional technical requirements for appearing in the AI Overview — the signal for the AI Overview is the same signal that determines classic ranking. Query fan-out is specific to AI Mode (the optional AI-first search). For Google AI Overview, the goal is classic ranking with properly structured content.

Claude — uses Anthropic's infrastructure (a search layer through Claude-SearchBot, a fetch layer through Claude-User). Claude especially values academic and documentary style, with caution toward marketing claims. For Claude citation, documented primary sources and named authorship are decisive.

Gemini — uses the Google Search index and the Gemini model. The signal largely overlaps with Google AI Overview, but Gemini as a conversational agent supplements it with a better understanding of follow-up questions and contextual resolution.

Related articles in this cluster structure go deeper into platform specificity: Google AI Overviews and AI Mode — how to get cited (coming soon) covers Google-specific mechanics; ChatGPT and Perplexity site optimization (coming soon) covers conversational AI systems; AI optimization techniques and llms.txt (coming soon) covers technical implementation.

Authority: how AI systems assess trust

Authority in AI optimization does not mean the same thing as classic PageRank. Authority is the set of signals that show an AI system a source is worth citing: an entity mention in the knowledge graph, named authorship with a verifiable identity, citations in primary and secondary sources, presence on review platforms (G2, Capterra, Trustpilot in B2B, industry catalogs in other lines of business), and brand mentions in text without a direct link. The AI system aggregates these signals into an internal trust assessment — the higher the score, the greater the chance of citation.

Entity signals — knowledge graph, sameAs, Wikidata

Wikidata is a collaborative knowledge structure through which Google connects entities to the Knowledge Graph. A brand with a Wikidata item that has its properties filled in (founders, founding date, headquarters, areas of work, language, website, sameAs links) is in a better-defined position for AI citation than a brand without a Wikidata presence. A Wikidata item is not mandatory for ranking, but it is the cheapest route into the knowledge graphs that AI systems read.

CrunchBase is a specific reference point for agencies, technology firms, and B2B brands. CrunchBase is referenced directly in Google's entity signals and in some AI aggregators. Professional directories — Clutch, Sortlist, GoodFirms for agencies, G2 and Capterra for software firms — build an additional layer of entity signals.

All these sources are connected in the sameAs array in an Organization or Person schema. When an AI system performs entity resolution — the process of linking a reference in text to a canonical entity — the sameAs network shortens the path and increases the reliability of the resolution.

Named authorship and the Person schema

An article with a named author who has a Person schema, a biography, and sameAs links to LinkedIn, ORCID, and GitHub is significantly better positioned for AI citation than an article with no author or an anonymous author. An AI system generating an answer preferentially cites a source with a verifiable author — it treats anonymous content as a weaker signal.

The Person schema should contain: first and last name, jobTitle, a worksFor reference to the Organization schema, and a sameAs array with at least three external reference points. The ORCID identifier is an especially strong signal for authors in the academic or research space; LinkedIn is the universal reference.

An author with an active publication history — several articles on the same topic, citations from other sources, mentions in the media — builds author authority in the knowledge graph. That authority transfers to every article the author writes.

The citation graph — primary and secondary sources

Cited expertise is a stronger signal than self-proclaimed expertise. An article that demonstrably cites primary sources (official documentation, academic papers, direct research) is treated as a secondary source that has been verified. An article that cites only other secondary sources is treated as a tertiary source — weaker.

Source strength is hierarchical:

1. Primary sources — Google Search Central, developers.google.com, official provider documentation, academic papers (arXiv, peer-reviewed journals), legal texts, direct research.
2. Reputable industry secondary sources — Search Engine Land, Moz, the Ahrefs blog, the Semrush blog, Backlinko, publications by recognized experts (Aleyda Solis, Barry Schwartz).
3. General secondary sources — technical publications (MIT Technology Review, InfoQ, TechCrunch), industry press releases.
4. Tertiary sources — other blogs, aggregators, reposts without attribution.

An article in the prestige register cites from the first two levels. This guide draws on the arXiv academic paper (Aggarwal et al.), Google Search Central, the Wikipedia GEO entity, the Search Engine Land analysis, Aleyda Solis publications, and the direct official documentation of Perplexity and Anthropic. That citation mix is part of the article's own authority.

Presence on review platforms and digital PR

Presence on review platforms (G2, Capterra, Trustpilot for B2B software and services; industry catalogs for other lines of business) is a direct entity signal. AI systems generating a recommendation answer preferentially cite sources with high ratings on relevant platforms — this is especially true for ChatGPT Search when a user asks for a recommendation.

Digital PR — a mention in the media and in industry publications without a direct link — builds the brand as an entity. This is a long-term investment, not a one-off campaign. In the context of AI visibility, every mention of the name in an industry context becomes a micro-signal that the AI system links to the entity.

The practice of entity building follows a clear order: Wikidata item → industry profiles (CrunchBase, Clutch, G2) → LinkedIn company page → author profiles with sameAs links → digital PR through HARO responses and industry commentary → citations in relevant publications. This order reflects the relationship between cost and impact — the low-cost steps that are measurably effective first, only then the expensive steps with a long-term return.

Measuring AI visibility: what to track and with which tools

AI visibility is not measured through the same dashboard as classic SEO. Direct AI referral traffic is structurally small (often below 1% of total traffic) and hard to measure (AI systems do not send attribution in the referrer header). Instead, AI visibility is measured through indirect signals: citation count, share of voice, sentiment, and the context in which the brand is mentioned. The measurement framework relies on tools that periodically send queries to AI systems and record the answers.

Primary and secondary metrics

Primary AI-visibility metrics:

• Citation count — how many times the brand is cited in an AI system's answer to a set of target queries. Measured monthly, by sending a standardized set of queries (8 to 30 per topic) to each AI system and counting mentions.
• Share of voice — the percentage of brand mentions relative to all brands mentioned in the same answers. A direct comparative measure against the competition.
• Sentiment — the positive, neutral, or negative context in which the brand is mentioned. The data comes from a structured analysis of the answers.
• Contextual triggering — which types of queries lead to a brand mention. It reveals gaps (topics where the brand is not mentioned although it should be) and strengths.

Secondary metrics:

• AI referral traffic — traffic from chat.openai.com, perplexity.ai, gemini.google.com, and similar domains. Measured in GA4 or another analytics tool via source/medium filtering. The value is indicative, not definitive — it is structurally small.
• Branded-search lift — growth in searches specifically for the brand on Google. The signal is indirect: a user who saw the brand in an AI answer remembers it and later searches directly.
• Direct-traffic lift — growth in direct traffic (without a referrer) that may be AI-inspired.

Measurement tools — industry practice

Semrush AI Visibility Index is one of the leading tools for systematically tracking AI citation. The benchmark report analyzed 2,500 prompts across five industries (Finance, Digital Tech, Business, Fashion, Consumer Electronics); the broader Semrush AI Visibility platform operates over a base of more than 100 million prompts per month (source: Semrush AI Visibility Index). Industry analyses through 2025–2026 document significant volatility of cited sources from month to month — AI visibility is not stable in the short term and requires continuous tracking.

Ahrefs Brand Radar (part of the Ahrefs subscription) tracks brand mentions across the digital space, including AI aggregator and index systems. It is useful above all as a signal of digital PR and indirect mentions.

Manual prompt testing — systematically sending 8–30 queries to each AI system, recording the answers in a structured document, with a periodic running review. For agencies and small brands that cannot afford a Semrush subscription, manual testing is a valid approach. The key is consistency — the same queries are sent to the same systems at the same interval (at least monthly).

The test protocol Praxidea applies to its own content combines three approaches:

1. Baseline — before publishing a new article, 8 queries related to the article's topic are sent to five AI systems (ChatGPT, Perplexity, Claude, Gemini, Google AI Overview) and the pre-publication state is recorded.
2. 30/60/90 days — the same queries are repeated at 30, 60, and 90 days after publication. Citation count per system and share of voice are measured.
3. Quarterly analysis — every three months the trend is synthesized: which articles show growth, which stagnate, where the gaps in topic coverage are.

A realistic time horizon

The first measurable results of AI visibility appear in the range of 6 to 12 months from the publication of a high-quality article. Faster indexing (the first 2 to 4 weeks) can result in appearance in the Google AI Overview if the article already ranks classically, but stable citation in ChatGPT, Perplexity, and Claude takes time for crawlers to index, for the entity graph to be enriched, and for authority signals to accumulate. The horizon of an infrastructure investment is 12 to 36 months — this is the time over which the cumulative effect of a series of articles, entity building, and the citation network develops.

The difference from classic SEO is mild but significant. Classic SEO results also accumulate over a similar time, but with clearer short-term signals (position in the SERP, clicks, impressions). AI visibility has no equivalent of Google Search Console impressions — it must be actively measured through prompt testing, which means the feedback loop is longer and more expensive.

The cost of AI optimization — how it is determined

The cost of AI optimization depends on the state of the site, the depth of the architecture, the state of the content, the strength of the entity, and the target market. A site with a correct technical layer and a clear author structure requires less work than a site built from scratch. A site targeting Serbian search requires a different scope from a site competing in DACH or UK search, where entity signals and domain authority are considerably stricter.

The exact cost is defined by an introductory audit, which establishes the state of the site, the priorities for intervention, and the scope of work. Praxidea does not publish price lists or standard packages; every engagement is run to the measure of the client's specific goal. Details about the audit itself: SEO audit cost: 0, 300 or 1000 EUR.

Questions and answers

My site does not appear in AI answers — why?

A site fails to appear in AI answers for three typical reasons, each resolvable at the architecture, content, or authority layer. Architecture layer: AI crawlers (GPTBot, OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot) cannot access the site — blocked in robots.txt, poor internal linking, a missing llms.txt, or no basic Article schema. Content layer: the content exists but is not structured for extraction — sections do not open with answer capsules, there is no named author with a biography, primary sources with dates are missing. Authority layer: the brand has no entity graph (no Wikidata item, the sameAs links are empty or point to irrelevant platforms, there are no industry citations). Diagnosis starts at the first layer — only when the architecture is correct do content and authority have something to attach to.

What is the difference between GEO and AEO?

GEO (Generative Engine Optimization) is the umbrella term for the discipline that optimizes content and infrastructure for appearance in generative AI search systems — ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude. It covers the discipline as a whole. AEO (Answer Engine Optimization) is a narrower term that focuses on optimization for systems that generate direct answers — featured snippet, FAQ extracts, Google AI Overview, AI summaries on other platforms. In practice the terms are often used interchangeably, but GEO covers a broader scope (including entity ranking, cluster architecture, crawler control), while AEO focuses more narrowly on the format that is extracted directly as an answer. In this guide GEO is used as the primary term, AEO as a related one.

Is it better to hire an agency or a freelancer for AI optimization?

The choice depends on the complexity of the site, the scope of work, and the need for several disciplines at once. A freelancer suits a predictable scope focused on one or two disciplines (e.g., structured data + llms.txt setup for a smaller site). An agency suits a complex scope that requires several disciplines at once — technical SEO architecture, content production, entity building, measurement and reporting. AI optimization often requires precisely that multiplied approach: the architecture is coded, the content is written, the entity is built systematically. A site with a clearly defined technical task can get by with a freelancer engagement; a site with a holistic goal (AI visibility as a continuous discipline over 12–36 months) usually benefits from an agency engagement with a structured methodology.

How often should content be refreshed for AI visibility?

Evergreen content on stable topics is refreshed every 6 to 9 months unless a trigger event occurs — Google announces a major AI Overview update, Gemini adds a new feature, ChatGPT changes crawler behavior, Perplexity changes the source-ranking algorithm. Time-sensitive content (specific regulations, deadlines, year-by-year statistics) should be updated before the trigger date. The update-date field in structured data (dateModified) signals freshness to AI systems — an article without a visible last-modified date is treated as a less reliable source. The update need not be a full rewrite — changing the primary source, adding a new section, or revising specific data is enough.

Should classic SEO be abandoned entirely?

No. Classic SEO and AI search optimization are not substitutes — they are additive. The same site that ranks classically in Google more often becomes a candidate for citation in the Google AI Overview; the same site with clean architecture, quality content, and authority is read better in ChatGPT, Perplexity, and Claude as well. Classic SEO remains the dominant source of traffic in 2026 — AI referral traffic is structurally small (below 1% of the total), while Google organic traffic makes up 40 to 60% of total traffic for most websites. AI optimization is an enhancement that broadens the sources of visibility, not a replacement that would supplant Google ranking.

Does AI optimization deliver a measurable ROI in the first 6 months?

Direct ROI in the first 6 months is structurally limited. Direct AI referral traffic is small, and the indirect effects (branded-search lift, direct-traffic lift, brand mentions in AI answers) accumulate more slowly and are harder to attribute. Realistically: the first 6 months are measured as a leading indicator — citation count in a control set of prompts, share of voice, entity resolution in the Knowledge Graph. The real ROI appears in the range of 12 to 36 months, when the cumulative effect of architecture investments, content depth, and authority signals becomes material. An agency or a client that expects a 6-month measurable ROI from AI optimization is treating the discipline wrongly — AI optimization is an infrastructure investment, not a marketing campaign.

---

This guide is the first article in the AI visibility and GEO cluster on the Praxidea site. Related articles covering specific platforms, techniques, and measurement frameworks are published successively; internal links to the related articles are added upon publication.

Choosing an agency for GEO follows the same criteria as choosing an agency for classic SEO — method, transparency of reporting, an explicitly stated scope of mandate.

Contact.