SEO didn’t die. The reader changed.

I work on SEO, AEO and GEO at Beyond Himalaya Tech, an Australian AI engineering firm. Separately, I rebuilt Transit Education’s website from WordPress to Next.js. That rebuild is the example I can describe here, because its details are already part of my own work.

The practical question was not whether to replace SEO with a newer abbreviation. It was whether a student could find an answer, understand it and take the next step. I still start there. The extra question now is whether another system can read that answer without getting the wrong idea.

Three names, overlapping work

I use SEO for the work of making a site findable and useful in search. I use AEO, answer engine optimisation, for making an answer clear enough to stand on its own. GEO, generative engine optimisation, adds the question of how a source appears inside a generated response. Those are my working distinctions, not three separate engines I can switch on.

The GEO paper gives the last term a research framework: it studies source visibility in generated answers, rather than just a page’s position in a list of links. That difference is worth understanding even if the terminology keeps changing (Aggarwal et al., 2024).

Google says AI Overviews and AI Mode can use related searches across subtopics to construct an answer. It also says a supporting page must be indexed and eligible for a search snippet. There is no extra technical admission test for those AI features (Google Search Central, AI features, 2025).

My reading of that is simple. The answer has another reader before it reaches the person. The underlying page still needs to work.

A crawler has to reach the page first

OpenAI separates its search crawler, OAI-SearchBot, from GPTBot, which concerns potential training use. Its documentation says those permissions are independent. Allowing search access is not the same decision as allowing training access (OpenAI, n.d.).

Perplexity describes PerplexityBot as a crawler for surfacing and linking sites in search, not for training foundation models. It separately describes Perplexity-User as a user-triggered fetcher. Its guidance also covers firewall access, because a robots rule is only part of the access question (Perplexity, n.d.).

I would therefore check the intended crawler’s rules and the actual response before commissioning a new batch of AI-friendly copy. A page that my browser opens is not, by itself, proof that every relevant crawler can reach it. That is a check I want evidence for, not an assumption.

At Transit, the concrete work included single-hop redirects from old WordPress addresses and one canonical host. Those were migration decisions, not an AI feature. I would keep them even if nobody used the term GEO again.

Write something worth carrying across

Transit’s country guides have structured FAQs. Its blog posts have one clean set of structured data. I think of those as ways of organising an answer, not permission to write a vague page and hide the useful bits somewhere else.

For a country guide, I would put the actual question beside the answer and keep any conditions with it. If a requirement depends on an institution or a course, that qualification belongs in the same paragraph. A short sentence that loses the condition is worse than a longer one that stays accurate.

Google’s people-first guidance asks whether content has clear sourcing, original value and an identifiable creator. It encourages accurate authorship information rather than invented credentials. That is a more useful starting point for authority than trying to make every paragraph sound confident (Google Search Central, Helpful content, 2026).

I want a reader to be able to tell what comes from my implementation, what comes from an official source, and what is my recommendation. I would rather leave an honest limit in the answer than make it easier to quote something misleading.

Research is useful, promises are not

Aggarwal and colleagues tested changes including citations, quotations and statistics. Some improved visibility in their benchmark, and the effects varied by domain. The paper also included experiments with the deployed Perplexity system of that period. It is evidence about those experiments, not a forecast for a restaurant or consultancy in 2026 (Aggarwal et al., 2024).

My takeaway is to make evidence easier to inspect. It is not to insert a statistic into every paragraph. If I do not have a number, I do not have a number. Decorating a claim with a citation that does not support it would defeat the point.

I generated llms.txt for Transit too. The current proposal describes a Markdown guide to information for agents. It remains a proposal, not a universal search-engine standard or a guarantee of citation (Howard, 2026).

Google is unusually explicit here: its current guidance says llms.txt neither helps nor harms Google Search visibility or rankings. I treat the file as an optional aid for systems that use it, not as a Google ranking task (Google Search Central, AI optimisation, 2026).

Separate being cited from being useful

Microsoft’s February 2026 AI Performance announcement describes citation counts, cited URLs and a sample of grounding queries across supported Microsoft AI experiences. It explicitly says those figures do not establish ranking or placement inside an answer. It is not a report on every answer engine (Microsoft Bing, 2026).

That gives me a sensible reporting boundary. I would record where an answer appeared, which page it cited and whether the answer represented the page correctly. Then I would keep visits and meaningful actions separate. A citation is an observation, not an enquiry or a sale.

I am not attaching an AI traffic increase to Transit here. The work I can point to is the rebuild, its generated search files, its structured content and the small SEO check script I added after catching repeated brand names in titles and missing sitemap pages.

That is where I would spend the next hour too: check a real answer, fix a real ambiguity, and make sure the next edit does not break the basics. Build less, think deeper makes sense to me when it means doing that work before adding another optimisation layer.

Sources

Documentation checked on 30 October 2026. Pages without a published date are marked n.d.

Google Search Central (2025). AI features and your website. Updated 10 December 2025. https://developers.google.com/search/docs/appearance/ai-features

OpenAI (n.d.). Overview of OpenAI crawlers. https://developers.openai.com/api/docs/bots

Perplexity (n.d.). Perplexity crawlers. https://docs.perplexity.ai/docs/resources/perplexity-crawlers

Google Search Central (2026). Creating helpful, reliable, people-first content. Updated 5 October 2026. https://developers.google.com/search/docs/fundamentals/creating-helpful-content

Aggarwal et al. (2024). GEO: Generative Engine Optimization. KDD 2024; version revised 28 June 2024. https://arxiv.org/abs/2311.09735

Howard (2026). The /llms.txt file, v2. First published 3 September 2024; updated 10 August 2026. https://llmstxt.org/

Google Search Central (2026). Google’s guide to optimising for generative AI features on Google Search. Updated 10 July 2026. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide

Microsoft Bing (2026). Introducing AI Performance in Bing Webmaster Tools Public Preview. Published 10 February 2026. https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/