SEO for artificial intelligence: how to get into ChatGPT and Google AI answers

Short answer: start with a page that makes clear what you do, who it is for and what supports your claims. Check whether search engines can read it, then measure what actually appears. No setting guarantees that ChatGPT or another AI system will recommend you. Below, I separate documented platform rules from a practical workflow I suggest for a small business.

SEO for artificial intelligence, often called GEO or AEO, is not one channel. A company mention, a link to your page, a visit and an enquiry are four different outcomes. For a custom application provider, I care most about the last one: did a potential customer understand the solution to their problem and take the next step?

What individual platforms actually document

ChatGPT: search is not model training

OAI-SearchBot supports search; GPTBot collects content potentially used for training. Their permissions are independent. ChatGPT-User makes user initiated visits; robots.txt may not apply, and it does not control search exclusion. Check the first for search access and the second for training preferences. OpenAI's official crawler documentation also publishes IP information for verification.

Google AI Overviews and AI Mode

A page must be indexed and eligible for a snippet to qualify as a supporting link. Inclusion is not guaranteed. Google may search related subquestions, without specifying a fixed query count. See its official AI feature requirements.

Perplexity

PerplexityBot supports search results; Perplexity-User makes user requested visits. The provider publishes IP ranges to verify access. Perplexity's documentation is the starting point for firewall and crawler rules, rather than assumptions based on what another search engine allows.

My practical conclusion: keep a separate record for each channel. Allowed access, an actual crawler visit and a citation in an answer are different pieces of evidence. If you also want to track Copilot, Gemini or Grok, add separate tests instead of treating your ChatGPT result as theirs.

What a small business should fix first

Before commissioning another article, I would review the main sales page. Imagine a business trying to stop copying orders from email into Excel. A broad promise about a digital future does not tell them whether you are the right provider. A process example does: where an order arrives, who approves it and where the information ends up.

1. Explain the problem and the limits of the solution. Name the audience, what the application handles and what a person still does. Distinguish building a new application from configuring an existing tool, and initial delivery from maintenance.
2. Show a verifiable example. For a public website, include its link and your actual scope of work. For an internal application, use a clearly labelled demonstration with fictional data or an approved scope description. Do not expose private screens or customer records for SEO.
3. Explain price alongside scope. If you publish a starting price, state what it includes, what changes it and what costs extra. If a figure cannot yet be determined, list the information needed for a quote. That is more useful than a precise looking number without conditions.
4. Resolve contradictions. Check the company name, contact details, services and actual coverage on your site and profiles you control. Do not invent branches or endorsements. When another site has incorrect information, request a factual correction rather than artificial promotion.
5. Connect the answer to the next step. An article about an ordering portal should lead to the relevant service or example, not only the homepage. Tell readers what information to send so they can receive a useful response.

Google's helpful content guidance emphasises original information, clear sourcing and identifiable authorship. That is not a formula for guaranteed citations. It is a useful editorial check: what can a reader verify here that a generic description of the service does not provide?

The technical check: what a system actually receives

Google can execute JavaScript; not every other crawler can. Server rendering or prerendering can reduce dependence on scripts executing successfully. Google's JavaScript SEO guide explains the distinction. A JavaScript website is not automatically invisible to search.

For an audit, I would save the initial HTML and compare the rendered page. Do they have the same title, main content, language and canonical URL? Does important information appear without a click, login or dialog? I would test direct entry to an individual page, not just navigation from the homepage.

• Check the page response, redirects and indexing instructions.
• Compare the canonical URL in the initial HTML with the rendered version.
• Check the actual Slovenian and English text, not just language tags.
• Open links to important pages and test the contact journey on a phone.
• When a crawler is blocked, investigate hosting and firewall rules too. Do not disable security controls indiscriminately.

A successful technical test means you have removed the confirmed obstacle. It does not mean the page is already indexed or will be selected for an AI answer. I would always report those states separately.

What not to buy as a guaranteed shortcut

• An llms.txt guarantee. Google Search ignores this file for visibility and ranking. That is not a claim about every AI system. Google also does not require splitting content into tiny chunks. Its generative search guide explains both limits.
• Special AI schema. Google requires no special structured data for AI features. Its official guidance does not say all structured data is useless.
• A score without a method. Before paying for a tool, ask for its question set, supported languages, sample dates and definition of a mention. A proprietary metric can help comparisons, but a percentage does not make it an official search engine score.
• A promise of first place or a fixed citation count. Ask for verifiable deliverables instead: which pages will change, which obstacles will be removed and how results will be tracked. A provider controls their work, not another system's choice of answer.

Measure outcomes without mistaking a bot for a customer

Google's Generative AI report in Search Console shows impressions in AI Overviews and AI Mode. It does not cover every AI tool. Insufficient impressions can be one reason the report is missing.

In web analytics, track recognised referrals and enquiries. Direct traffic is not evidence of an AI visit: Google Analytics explains that it lacks a clear referral source, with several possible causes. Do not attribute all direct traffic to ChatGPT.

I suggest three separate records. The first captures technical access and blocks. The second saves the question, full answer and cited link. The third records business outcomes: an identifiable visit, an enquiry and whether it fits your service. A crawler visit belongs in the first record, not the third.

For each comparison, record equal time periods, language, country, the tool used and whether web search was enabled. Keep a dated change log for your site. If you change the offer, headline and measurement method together, you cannot honestly attribute a later movement to one of them.

Test it yourself in 20 minutes

Open the available search modes in ChatGPT, Perplexity and Google. Start a fresh conversation and use the same questions a customer would ask. Replace the brackets with your service and region:

1. Who offers [your service] in [your city or region]?
2. How much does [your service] cost in [your country]?
3. What should I look out for when choosing a [your profession]?
4. [Your company name], what do they do?
5. Compare providers of [your service] in [your region].

This is an initial sample, not a scientific assessment of the whole market. Read the results like this:

• No mention: you were not selected for this answer. Inspect the sources and repeat the test; this does not prove your company is unknown to the system.
• Wrong description: open the cited source and check the fact. The error could be in the source or the answer, not necessarily on your site.
• Competitors appear: record who, with which link. Then check whether you address the same need; the result alone does not explain the difference.
• A generic answer without names: the question may not have triggered a provider comparison. Do not infer that the market is empty.

Repeat the test on an agreed schedule and retain unsuccessful results too. For an initial review, I would choose weekly checks and a monthly review of enquiries. That is a suggested working rhythm, not a deadline by which visibility can be promised.

What the audit should actually deliver

A useful report should identify the confirmed problem, the affected page, the proposed fix and how to verify it. The next task might be a clearer description of an application's scope, a corrected translation or a removed block, rather than another article.

For a website or custom application, my suggestion is to start with one real business problem and one clear path to an enquiry. Judge improvements by evidence and usefulness to the customer. The goal is not a prettier SEO score. It is less uncertainty between your offer and the person who needs it.

Official sources were checked on 20 September 2026. Links appear beside the claims they support; the suggested workflow is not a ranking guarantee.