AI Search: Frequently Asked Questions

The primary AI systems to optimise for in 2024-2025 are Google AI Overviews (which appear in Google search results), ChatGPT (the most widely used standalone AI assistant globally), Perplexity AI (popular among research-oriented users), and Microsoft Copilot (integrated into Bing and Microsoft 365). The good news is that the GEO principles that help one system recognise your authority generally help all of them — the underlying criteria for trustworthiness and citability are consistent across platforms. Rather than chasing each platform separately, focus on the shared fundamentals — accuracy, structure and authority — that earn citations everywhere.

Test your AI visibility by opening ChatGPT, Perplexity, or Google AI Overviews and asking: "What is the best [your service type] in [your city]?" or "Who are the leading [your industry] experts in South Africa?" Note whether your business name appears, how it is described, and what sources are cited. Repeat this test monthly as you implement your GEO strategy to track improvement. Google Search Console also provides data on whether your pages are appearing in AI Overview results. Keep a simple log of the answers you see each month so you can measure progress over time.

Measure it in your own analytics rather than relying on spot-checking AI answers. In Google Analytics 4, traffic from ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews arrives as referrals from those platforms' domains — commonly grouped and reported as "GEO sessions". Four numbers tell you most of what you need: sessions and users from AI platforms (is anyone arriving at all), engagement rate and average engagement time (is the traffic qualified), key events (is it doing anything), and purchases or leads (is it paying). The most useful starting point is a zero baseline: if the site currently receives no AI-platform sessions, every session that appears afterwards is unambiguously incremental, with no prior traffic to reattribute and no seasonality to argue about. Pair that with a monthly log of how AI assistants describe your business, since citations without clicks are real visibility that analytics alone will not show you.

In the GEO programmes we have measured, yes — and those visitors engage rather than bounce, which is what you would expect from someone who has effectively been pre-qualified by the answer that sent them. Across our anonymised case studies: a safety training provider recorded 46 purchases worth over A$9,000 from AI-platform traffic in four months; a cooking school recorded 9 tracked purchases; a vending provider saw average engagement time rise from 28 seconds to 1 minute 17; and a material-handling supplier saw a 69.11% engagement rate from AI-referred visitors. The volumes are smaller than traditional organic search today, so this is not a replacement for SEO — and because we do not have channel-by-channel comparisons for these accounts, we are not claiming AI traffic converts better than organic, only that it converts at all. What makes it worth the work is that it is growing, and that in two of those cases it started from a baseline of literally zero — meaning every sale was additional to what the existing SEO work was already producing.

Not any more, provided you choose the right setting. Until 15 September 2026, telling Cloudflare to block AI training also blocked mixed-use crawlers — Googlebot, Bingbot and Applebot all do more than one job, so a rule aimed at model training removed ordinary search indexing as well. On 15 September 2026 Cloudflare introduced a fourth option, Disallow AI Training, which publishes a no-training preference into your robots.txt while leaving those mixed-use crawlers allowed for search. The four settings are now Allow, Disallow AI Training, Block on pages with ads, and Block. The last two still stop Googlebot doing either job. Cloudflare's own figures suggest the separation is what most people wanted: 17% of its sites block training in some form, while under 1% choose to block Search bots. If you set this before 15 September 2026, it is worth going back and checking which option you are actually on.

Most likely because of a logging fault at Google rather than anything you did. A logging error caused sharp, artificial declines in the generative AI performance report between roughly 12 and 20 August 2026, and it was corrected around the 20th. If your chart shows a dip in that window, check the dates before investigating your content or your configuration. Any month-on-month comparison spanning those days is comparing against a fault, not a performance change. It is also worth remembering what the report measures: impressions on AI surfaces, already counted inside your existing Search totals rather than added to them. It tells you that a page was shown. It does not tell you that you were cited, quoted, recommended or clicked.

Considerably less than it did. Reddit's share of ChatGPT citations fell from roughly 3.8% to about 0.5% over four days in August 2026, an 86% decline, following a change at OpenAI. The cause has not been fully established publicly, so treat any confident explanation you read with caution. It is a large enough drop to change the economics of any strategy built primarily on seeding Reddit threads. This is worth understanding as a pattern rather than a one-off: when a single source becomes a reliable route into AI answers, it attracts manipulation, and the platforms then downweight it. Building your visibility on whichever third-party surface is currently cheap to influence means rebuilding it every time one of these corrections happens. Participating in communities where your customers genuinely are remains sensible; treating any single platform as a citation channel is not.

Often not, and you should measure your own site rather than trust a general figure. Google renders JavaScript, though on its own schedule and not instantly. Several AI crawlers do not render at all — they read the HTML your server sends and nothing more. You will find confident percentages circulating about exactly how much each crawler executes; we do not quote them, because the numbers in circulation are mostly secondary citations of each other rather than traceable studies. The test that actually answers the question for you takes an hour: fetch an important page without executing scripts, compare the links in that raw HTML against the links your browser shows, and then check your server logs to see which URLs each crawler has genuinely requested. Links matter more than body copy here — a link that only exists after rendering is a page some crawlers can never discover.

When an assistant answers "who should I use for this?", it compares the evidence it can find for each candidate: how clearly each business is identified, whether trusted third parties mention it, and whether its own pages state specific, checkable facts. A competitor appearing usually means more of that evidence is findable, not that they are better at the work. The useful response is a comparison rather than a guess. Run the same ten buyer prompts across several assistants, record who appears and which sources are cited, then close the specific gaps: entity details, third-party mentions, answer-first pages. Results differ by assistant and shift from week to week, and nobody can promise you a placement.

Because the assistants are not drawing on the same sources. In an Orbit Media study of 13,184 citations across 1,765 answers (72 prompts, three B2B brands, a directional sample the author describes as not statistical proof), ChatGPT, Claude, Perplexity and Gemini all cited the same domain for the same question in just 30 of 1,792 query-and-domain combinations, about 1.7%. Perplexity cited an average of 19.2 sources per answer against 4.5 for ChatGPT, and leaned on LinkedIn, YouTube and Reddit where ChatGPT favoured primary documentation. Being visible in one assistant therefore says very little about the others. Measure each platform separately, with the same prompts each month.

It depends on who measured, so quote a range and name the source. Published 2026 estimates include 43% (Similarweb, reported 30 July), over 62% of queries with a range of 55% to 84% by industry (seoClarity), a blended "over 80%" that Search Engine Roundtable reported on 26 October from Semrush, SE Ranking and seoClarity data (the only per-source figures in that piece are SE Ranking at 82.5% of US brand searches and seoClarity at over 62%) and 86.7% (Peec AI, a vendor, from 500,000 prompts in April). The studies sample different keywords, countries and dates, and the share has been rising. For South Africa, Algorithm checked 5,850 non-branded page-one terms for its clients and found an AI Overview on 4,100 of them, 70%, in a single snapshot. The number that matters most is the share on your own 30 to 50 key searches.

Not the fundamentals, but they raise the cost of getting them wrong. An assistant comparing retailers may struggle with a price assembled by a script, a stock status that is only implied, or a delivery time buried in a footer. Put the price, availability, delivery timeframe and returns terms in plain text on a stable URL, repeat them in Product and Offer markup that matches the page, and test that checkout works without a puzzle only a person could solve. We are not aware of any published measurement showing that this changes which retailer an agent chooses, so treat it as removing ambiguity, not as a recommendation lever.