How to do GEO, in the right order.

Why GEO matters now

AI search query share crossed a threshold in 2026. Standalone AI engines (ChatGPT, Perplexity, Claude, Gemini, Copilot) plus Google AI Overviews and AI Mode now intercept a meaningful slice of buyer research that previously hit traditional Google results. Pew Research found users clicked a result in 8% of visits with an AI summary against 15% without. MailOnline reported CTR 56.1% lower on desktop and 48.2% lower on mobile. Ahrefs found an AI Overview correlates with a 58% lower click-through rate for the top-ranking page (December 2025 data). Seer measured a 61% fall in organic CTR on AI Overview queries to September 2025, then a partial rebound in January and February 2026. Brands either get cited inside the AI answer or get cut out of the consideration set entirely. That is the practical case for running GEO as a discipline, not a side project. The five stages below sit inside the longer GEO field guide if you want the wider context first.

Stage 1: Baseline

Run your brand and three to five named competitors through each major AI engine on the 10-20 buyer queries that matter. Capture what shows up. The free AI Search Visibility tool scores structural readiness and gives you test queries to run in ChatGPT, Perplexity and Google AI Mode yourself; use a paid tracker (Profound, AIClicks, Otterly) for ongoing monitoring. The baseline tells you which engine is the actual gap (often it is not the one you assumed) and which queries are open vs locked. Skipping this stage is the single most common mistake - teams jump to fixes without knowing where the gap actually sits.

Output of this stage: a citation matrix - rows are priority queries, columns are engines (ChatGPT, AI Overviews, Perplexity, Claude minimum), cells show present/absent and the cited domains. This becomes the reference document the rest of the programme works from.

Stage 2: Foundations (the SEO-overlap layer)

Bot access verified for the search crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot), with GPTBot, ClaudeBot and Google-Extended set as deliberate training choices. Google-Extended also controls grounding in the Gemini app, so blocking it also affects Gemini app answers. Schema chain in place - Article, Person, Organization, FAQPage where the page answers explicit questions. Named author on every content piece with proper credentials. Content depth on priority topics (1,500-2,500 words on the queries that matter). This stage is recognisable as SEO; it is the work that AI engines and Google both reward.

Time cost: 30-90 days depending on existing SEO maturity. If foundations are weak, this is where most of the early budget goes. If foundations are strong, skip to Stage 3 and inherit the wins.

Stage 3: Citation surface (engine-specific)

This is where you target the gap the baseline revealed. OAI-SearchBot access for ChatGPT, and Bing index health for Copilot. PerplexityBot access for Perplexity; Ahrefs' data does not show Perplexity favouring fresh content more than other assistants, and its in-text citations averaged 1,166 days old. Knowledge Graph and YouTube transcripts for Gemini. Ranking and content quality for AI Overviews, not more schema (see the Ahrefs test below). Reddit presence where your audience lives (Reddit was the most-mentioned domain in ChatGPT answers in Ahrefs' September 2026 US data, at 16.8%). Do not do every engine - prioritise based on which one your buyers actually use.

Important: per Ahrefs' May 2026 test on 1,885 pages, adding JSON-LD gave no citation lift in AI Mode or ChatGPT, and a small decline in AI Overviews - it is hygiene rather than a standalone lever. The ranking position and the named-author signal do most of the work. Do not get stuck adding more schema thinking it will fix a content-depth problem.

Primary sources for the GEO stack: GEO paper by Princeton, IIT Delhi and other researchers (Aggarwal et al., 2023, KDD 2024), Google's guide to optimising for generative AI features (May 2026, updated July 2026), Search Engine Land's AI crawler optimisation guide, Profound (citation tracking).

Stage 4: Measurement

Weekly or fortnightly sampling against your priority queries. Track three metrics: citation frequency (how often you appear), citation share-of-voice (your appearances vs named competitors), and query coverage (how many of your priority queries cite you at all). Daily creates noise because AI outputs vary per session. Monthly is too slow if you are actively running work.

What to ignore: raw retrieval signals (your page appears in the model's context but is never named), branded queries (vanity, no signal about category authority), and traffic numbers from GA4 alone (AI engines often cite without a click - your brand mention is the win, not the visit).

Stage 5: Iteration

Quarterly content refresh based on the measurement data. Identify queries where you appear and queries where you do not. Where the gap is content-shaped (a competitor with more depth gets cited), add depth. Where the gap is extraction-shaped (your content is there but ChatGPT picks a cleaner source), restructure for extraction. Where the gap is authority-shaped, work the citation pool slowly (Wikipedia entry, Reddit presence, trade publication mentions - none of these compound in a month).

How to spot a GEO programme that is failing

Three patterns I see repeatedly:

• Six weeks of work then silence. GEO is an ongoing discipline. Burst-then-quiet does not build citation density. If the programme has a "campaign" shape, it is not actually GEO.
• "We added schema everywhere." Schema as the headline action means the consultant did not run a baseline. There was no diagnostic - just hygiene work badged as strategy.
• No named-author work on the agency's own site. If they have not done Person schema + credible bios + sustained Reddit presence on their own brand, they cannot do it on yours. Check before you sign.

Common mistakes

• Skipping the baseline. Teams optimise for ChatGPT when the gap is actually in AI Overviews. Wasted budget.
• Doing every engine. Spread thin across six engines, deep on none. Pick the two or three that match your buyer.
• Treating GEO as a campaign. It is an ongoing discipline. Six weeks of work followed by silence does not build citation density.
• Confusing retrieval with citation. ChatGPT pulls in pages it never cites: in Ahrefs' April 2026 study, 67.8% of the URLs it retrieved but did not cite were from Reddit. Yet Reddit was also the most-mentioned domain in ChatGPT answers in Ahrefs' September 2026 US data, at 16.8%. Track citation, not retrieval-adjacent metrics.
• Outsourcing community presence. Reddit accounts that look fake get downranked. The named expert on your team has to do the posting.