Does GEO Work? What the Evidence Shows

Does generative engine optimization work? It does in a narrow, measurable sense, and most of the bigger promises made in its name are still unproven. Published research shows that rewriting a page can change how much an AI-generated answer draws from it. The companies that run AI search tell site owners to allow their crawlers and publish useful content, and their published guidance describes no special trick. As far as we can find, no one has yet published a controlled test showing that GEO work made an AI assistant recommend a local business more often.

If you want the basics first, our guide to generative engine optimization explains the term. In one sentence, GEO is the work of being named and described correctly when someone asks an AI assistant like ChatGPT or Perplexity a question. This post covers what the research found, what the platforms say, what we have tracked for one of our own clients, and how to judge a GEO pitch before you pay for one.

What Did the Original GEO Research Find?

The term comes from a research paper, GEO: Generative Engine Optimization, first posted in November 2023 and published at the KDD 2024 conference. The authors built a benchmark of 10,000 queries and tested different ways of rewriting the web pages an AI engine used to answer them. They measured how much of each answer came from each page and how early that material appeared. Their headline finding was that GEO methods could boost visibility by up to 40% in generative engine responses.

The methods that worked best were ordinary editing choices. Citing sources, adding quotations and adding statistics produced a relative improvement of 30 to 40% on the paper's word-count visibility measure. Keyword stuffing, the old habit of repeating search terms, offered little to no improvement. Lower-ranked pages gained the most: citing sources raised visibility 115.1% for pages ranked fifth, while the top-ranked page lost 30.3% on average.

Pages that backed up their claims with sources, numbers and quotes earned more space in the answers, which points to careful writing more than to a new kind of service.

Key Takeaway: The founding GEO study found that sourced, specific writing earned more space in AI answers, while keyword stuffing did almost nothing.

Where the Research Stops Short

The paper is careful work, but its setup is a long way from a homeowner asking ChatGPT for a plumber. The main test engine fetched only the top five Google results for each query before writing an answer. A page had to be in those five before any rewriting could help it, so GEO in this experiment worked on top of search rankings.

The live test had limits too. The researchers ran their methods on Perplexity, but Perplexity does not let users pick source URLs, so they uploaded the source text as files. That shows how the model treated text it was handed, and it does not show how Perplexity finds and chooses pages on the open web.

The paper's visibility measures count words in an answer and score its impression, and neither tracks phone calls, booked jobs or whether the AI recommended a business by name. The authors also note that their methods may need to adapt over time as generative engines evolve. Anyone quoting the 40% figure as a forecast for a local contractor is stretching a lab result past what it measured.

Key Takeaway: The strongest GEO research measured pages that already ranked, on older AI systems, in words rather than calls, so it cannot predict what a GEO project will do for your phones.

What AI Search Tools Tend to Cite

A 2025 study from the University of Toronto, Generative Engine Optimization: How to Dominate AI Search, compared the sources AI search tools cite with the results Google returns. It found that AI search shows a systematic and overwhelming bias toward earned media, meaning third-party sources, over brand-owned and social content. In U.S. automotive searches, 81.9% of AI search sources were earned media, compared with 45.1% for Google.

The study also ran local searches, which is where it comes closest to home service. The researchers sent queries such as "best home cleaning services near me" to Google and to a web-enabled AI engine, then measured how many sites appeared in both. The overlap was low, at 20.6% for home cleaning and 17.1% for roofing. The authors say AI engines are less aligned with Google in surfacing local service providers, possibly because they weigh local authority and business directories differently.

A September 2026 preprint, Evaluating Brand Retrieval and Ranking in Large Language Model Recommendations, adds a warning about consistency. Its authors note that repeated responses to the same query can produce different brands and rankings. That study covered six AI models and five product categories rather than contractors, so treat it as a pattern to watch.

Key Takeaway: AI tools lean on third-party sources, their local picks often differ from Google's, and the same question can return different names, so one answer on one day proves very little.

What the Platforms Themselves Say

The companies behind these tools publish guidance, and it is more modest than most GEO pitches. Google's guide to succeeding in its generative AI features says the best practices for SEO continue to be relevant because its AI features are rooted in its core Search ranking and quality systems. It also says you don't need new machine-readable files, AI text files, markup or Markdown, and that structured data isn't required for generative AI search. From Google Search's point of view, the guide says, optimizing for generative AI search is still SEO.

Google's AI Overviews and AI Mode fall under what we call answer engine optimization. The chat assistants outside Google say less, and what they say is about access. OpenAI's crawler documentation says OAI-SearchBot is used to surface websites in ChatGPT's search features. It recommends allowing that crawler in your robots.txt file, which tells crawlers what they may visit, and says sites that opt out will not be shown in ChatGPT search answers, though they can still appear as navigational links. Perplexity's bot documentation gives the same advice for PerplexityBot.

Some platforms now report AI data back to site owners. In February 2026, Microsoft introduced AI Performance in Bing Webmaster Tools as a public preview showing how often a site's content is cited in Copilot, AI summaries in Bing and select partner tools. Google Search Console now has a generative AI performance report with impression data for AI Overviews and AI Mode, which Google says rolled out to all websites worldwide as of August 31, 2026. OpenAI's and Perplexity's crawler pages describe no similar report for site owners.

Key Takeaway: The platform documents describe no GEO trick; they ask for crawler access and useful content, and only some tools report AI data back to site owners.

Our AEO, GEO and SEO Work With Lackawanna College

We work with Lackawanna College, a Scranton college serving the people of Northeastern Pennsylvania, on AEO, GEO and SEO. Our work focuses heavily on the college's trades programs, which help people get into the trades through education, including its trades centers in Bethlehem, Chambersburg and Greensburg. The numbers below come from Semrush AI Visibility, a third-party tool that tracks how AI engines describe a brand. They measure the college as a whole, not the trades programs alone, and the trades centers opened in June and July 2026, after our baseline.

A college is not an HVAC or plumbing company, so read these numbers as one example of how AI answers about a business can be tracked over time. Here are the four snapshots, from the May 1 baseline to the September 29 data.

• May 1, 2026 baseline, Google AI Mode: AI answers that mentioned Lackawanna were 75% favorable and 25% general, and the favorable share had dropped 7 points while all four tracked competitors moved up. Cost and financial support was the weakest topic at 53% favorable, campus experience sat at 67%, and the college's own website was cited only twice across the tracked questions.
• May 31, 2026, same tool and settings: Favorable sentiment had reached 100%, and cost and campus experience were each 100% favorable where they came up.
• June 28, 2026, Google AI Mode: Favorable sentiment held at 100%.
• September 29, 2026, Google AI Mode: Favorable sentiment was still 100%, though that period had only 2 AI mentions.

Reach has not moved the same way. The 100% figure describes the tone of answers that mention Lackawanna, and those answers are still few. The share-of-voice figures in this post are for Google AI Mode only and do not cover other AI engines. In Google AI Mode, share of voice in broad, unbranded answers sat near 0% in the May 31 and June 28 data, and in the September 29 data, measured against a different set of competing colleges, it was 0.18%.

This is one client, one tracking tool and no control group, so it does not prove GEO work caused the change. It shows how AI answers about the college moved over the period of our work. The next phase is growing the reach and impressions of the college's programs. Our GEO page explains what we can and cannot measure.

Key Takeaway: At Lackawanna College, favorable AI sentiment rose from 75% to 100% between May 1 and May 31, 2026, and was still 100% in the September 29 Google AI Mode data, on only 2 mentions. Share of voice in Google AI Mode is still near zero, and growing it is the next phase of our work.

How to Judge a GEO Pitch

Most GEO pitches sound confident, and the evidence above gives you a way to test that confidence. These are the claims worth pushing back on before you sign anything.

• A guaranteed mention inside ChatGPT or another assistant. None of the platform documents above describes a way to guarantee one, and answers can change between runs of the same question.
• A special AI file or markup sold as the way in. Google says you don't need new AI text files or special markup to appear in its Search, and the OpenAI and Perplexity pages cited above ask for crawler access.
• Seeded mentions on forums or review sites. Google says seeking inauthentic mentions across the web isn't as helpful as it might seem. Our breakdown of real SEO versus fake SEO covers how shortcuts like this tend to end.
• One screenshot offered as proof. Ask for the full question list, the dates, the tools checked and the misses.
• The 40% research figure quoted as a forecast. That number came from a lab benchmark in which pages already sat in Google's top five.
• A separate AI package priced on top of SEO and local work. Ask which tasks differ from accurate listings, review work and pages that answer real questions.

A pitch worth hearing starts with the basics. It covers a complete Business Profile, one name, address and phone number used everywhere, a steady flow of recent reviews, a site that lets AI search crawlers in, and service pages that answer the questions homeowners ask. Our guide to NAP consistency explains why matching listings carry so much weight in local search. A good pitch also names what can be measured, admits what cannot, and tells you before you pay how you would know the work was not paying off.

Key Takeaway: A trustworthy GEO pitch rests on listings, reviews and useful pages, and it explains its measurement limits before it asks for money.

What GEO Means for a Home Service Owner

GEO has research behind a narrow claim: specific, well-sourced pages can earn more space in AI answers once an engine has found them. The platforms point owners back to crawler access and useful content, and Google's guide calls optimizing for its generative AI features SEO. As far as we can find, no published controlled test yet shows that a GEO project sold on its own will get your company recommended by ChatGPT or Perplexity more often.

For an HVAC, plumbing or roofing company, it makes sense to fund work that pays off in search and also gives the AI search crawlers described above accurate, useful pages to read. That means accurate listings, recent reviews and pages written from real job experience, kept current every month. Google says businesses with complete and accurate info are more likely to show up in local search results, and that more reviews and positive ratings can help local ranking. Our SEO work and GEO work run on that same foundation, so nobody should charge you twice for the same tasks under two names.