We don't try to game AI answers. Here's what we measure instead.

Some of the advice now sold as "GEO" aims to get an AI assistant to name you whether or not you're the best answer to the buyer's question. We don't do that work, for clients or for ourselves. This post says why, in Google's own words, and explains why the proof usually offered for it, a screenshot of a brand at the top of a ChatGPT answer, doesn't measure anything.

We read every source below on 1 October 2026.

Google has put it in writing

Google's spam policies, last updated 28 August 2026, describe spam as techniques used to deceive users or manipulate Google's systems, "such as attempting to manipulate Search systems into ranking content highly or attempting to manipulate generative AI responses in Google Search." The last clause covers AI Overviews and AI Mode by name.

The same page lists "Creating multiple sites with the intent of hiding the scaled nature of the content" as scaled content abuse, and says Google finds violations through "automated systems and, as needed, human review." ChatGPT, Claude and Perplexity run their own search layers. Whatever their rules say, a result that depends on misleading the answer can vanish the next time the model or its sources change.

What it looks like this year

We've seen three versions of it this year, in advice aimed at small businesses and in the campaigns behind that advice. We won't name the people involved. The practices are the point.

• The coined term. Give an ordinary service a branded name, announce it in a paid press release, back it with a few videos, and build a site that credits you as the term's originator. When an AI answer explains the term, every source it can find is one you paid for or own.
• The satellite network. Build automated sites around your industry, or around an invented brand, so your name sits on more of the pages AI answers draw from. Google's scaled-content clause above describes this closely.
• The borrowed headline. Buy a placement on a news syndication network, then describe it elsewhere as editorial coverage. An answer that cites the second page inherits the description.

We stay out of all three for a plain reason. They change what the engines read without changing anything true about the business. An answer that names you because you wrote every source behind it isn't evidence that a buyer should choose you, and it rests on sources the engines, and Google's spam team, can re-weight at any time.

A screenshot isn't a measurement

The proof usually offered for these tactics is a screenshot: a brand at number one in a ChatGPT answer. Two things make that weaker than it looks.

First, the answer may be personal. OpenAI's help page on ChatGPT search says: "If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query." Its own example is a user who has said they're vegan and live in San Francisco, whose question about nearby restaurants becomes a search for vegan restaurants in San Francisco. A screenshot taken in a signed-in account can show what that account's history asked for.

Second, the answer moves. Detailed's AI volatility study checked more than 1,300 prompts once a day for four weeks, from the US with a non-logged-in account. Pick any two days and "ChatGPT returned an identical set of brands on 0.3% of pairs." The study also split each prompt's brands into a core, named on at least 80% of days, and a tail. On ChatGPT the core changed by 13% from one day to the next, against 78% for the tail. (The author notes that he works at Ahrefs, which sells prompt tracking.)

So one answer is a sample of one. A brand at the top of today's list may be in the tail tomorrow. The useful question is a rate: across repeated asks, how often does the engine name you, and who does it name instead?

What we measure instead

This is the method behind the SA AI Visibility Index and our client work.

• Four engines, ChatGPT, Claude, Gemini and Perplexity, each with its own live web search switched on.
• The same buyer questions every cycle, each asked several times, through each engine's API rather than anyone's personal account.
• Whether your brand is named in the answer, not only whether your site is cited, and which competitors are named instead.
• A published prompt panel and methodology, so anyone can check how a number was made.
• Rates rather than single answers, with the noise stated. Where we couldn't measure something, the report says so.

What we won't claim: that any tactic guarantees a recommendation, or that one engine stands in for "AI". Other tools now track parts of this too. What we sell is the method and the reading of it.

What does move an answer

The legitimate work is slower:

• Get onto the sources the engines already read in your category, such as the directories, review sites and guides, on your merits.
• Publish pages that answer the questions your buyers ask, not pages that list your offices.
• Let each engine's search crawler read your site. Our free crawler check tests OAI-SearchBot, Claude-SearchBot and PerplexityBot.
• Then measure again, the same way, and watch the rate.

Three questions for any "we got you into ChatGPT" claim

1. Was the answer taken signed out, and on how many engines?
2. How many times was the question asked, and how often did your brand come up?
3. Which sources did the answer cite, and who wrote them?

If the answers are "signed in", "once" and "we did", you've been shown a screenshot, not a result.

Sources, all read 1 October 2026: Google Search Central, Spam policies for Google web search (last updated 2026-08-28); OpenAI Help Center, ChatGPT search; Detailed, AI volatility (published 23 September 2026; prompts checked 26 August to 22 September 2026).