The Local AI Shortlist Report

There is already a version of Answer Engine Optimization that gets on my nerves.

Add an FAQ. Break every paragraph into tiny pieces. Create an llms.txt file. Publish a stack of articles written for a machine. Add some schema and call it AEO.

I don't think that gets to the heart of what is happening.

I started paying close attention to AEO after businesses I work on began appearing in AI answers in ways that were hard to dismiss as simple keyword matching. Sometimes the business was mentioned. Sometimes the answer understood what actually made the company different. In the strongest examples, the business landed on a short list of companies the system would consider hiring.¹²

Then I saw the other side of it. I could change the question and make the same business disappear.

People selling “AI rankings” tend to leave that part out.

The goal isn't to make an AI know your name. The goal is to make the short list when somebody asks who they should hire.

Based on my own testing, Google’s documentation, current AI search research, and the results I'm seeing across several properties, I think a business has to clear four different hurdles:

Found. Matched. Proven. Shortlisted.

Google hasn't published an “AI shortlist algorithm.” OpenAI hasn't either. These aren't four disclosed ranking factors. They're the framework I use at 580 Digital Infrastructure to make sense of what we can actually observe.³⁴

It explains quite a bit.

There probably isn't one AI ranking

Traditional search taught us to think in positions.

Search for custom showers OKC and somebody is first, somebody is second, and somebody is buried on page two.

AI search doesn't behave that cleanly.

Google says AI Mode and AI Overviews can use query fan-out. Its systems may run several related searches across different subtopics and sources before building an answer. Google also says AI Mode is designed for questions that involve comparison, exploration, and more complicated decisions.³

OpenAI describes a similar search behavior in ChatGPT. ChatGPT Search may rewrite a user’s question into one or more targeted searches and may use general location information to improve local results.⁴

These two questions may follow different retrieval paths:

Who has the best custom showers in OKC?

Who should I hire for a luxury walk-in tile shower in Oklahoma City?

A person may see those as nearly the same question. The system may not.

A 2026 preprint studying production retrieval-augmented recommendation systems found low overlap when the same buying intent was expressed through different natural paraphrases. Recommendation sets overlapped only 14 to 29 percent across paraphrases. When researchers reran the exact same prompt, overlap was 50 to 61 percent.⁵

Another 2026 preprint examined 3,750 responses across three models and found only 41.6 percent agreement on the top-recommended brand in its sample. A first-place recommendation on one model did not reliably remain first on another.⁶

I'd be skeptical of anybody selling an “AI ranking” as if there were one fixed position to track. The evidence doesn't support it.

The useful question is whether a business keeps entering the right consideration sets for the kinds of questions its customers ask.

1. Found

Before an AI system can recommend a business, it needs usable information about that business.

Google is unusually clear about this part. Pages that appear as supporting links in AI Mode and AI Overviews need to be indexed and eligible for Google Search. Google’s current guide describes retrieval-augmented generation as relying on core Search ranking systems to retrieve relevant, current pages from its index.³

The search foundation still matters.

Google now addresses AEO and GEO directly. Its position is that optimizing for generative search is still part of SEO because those experiences rely on Google Search infrastructure.³

I don't treat AEO as a replacement for SEO. If your pages can't be crawled, indexed, or retrieved, clever formatting won't save them.

For local businesses, Google includes accurate, current Business Profile information among the existing SEO practices that remain useful for visibility in its AI features.³

The first question is simple: does your business even enter the pool of information available for this search?

Plenty of businesses fail right there.

2. Matched

Being found isn't enough. The system has to decide that the business fits what the person asked for.

Google describes traditional local results through three broad ideas: relevance, distance, and prominence. Relevance measures how well a Business Profile matches the search. Distance considers where the searcher is. Prominence reflects how well known the business is.⁷

I wouldn't claim those are the complete ranking factors for Google AI Mode. Google hasn't said that. It has said its AI search features rely on core Search systems and can use local-business information, so the same basic questions remain:

What is this company? Where does it work? Does it actually fit the request?

Vague marketing copy becomes a problem here.

Imagine two contractor websites.

The first says:

We deliver quality craftsmanship and unmatched customer service.

That could describe almost any contractor in the country.

The second says the company builds custom tile showers in Oklahoma City, performs complete tear-outs, installs waterproofing systems, builds niches and benches, installs linear drains, and has completed more than 250 custom showers.

The second site gives an answer engine something useful.

I've watched this happen with Parsons Stone & Tile.

In one recorded ChatGPT test, I asked, “Who has the best custom showers in OKC?” I didn't mention Parsons in the prompt.¹

ChatGPT found the company on its own and described Parsons as its strongest candidate for the workmanship “behind the walls.” The answer specifically mentioned waterproofing, custom showers, niches, benches, linear drains, more than 250 showers of experience, and the Schluter system.

It went farther than recognition. Under a section about who it would contact for its own house, ChatGPT included Parsons among the three companies it would personally ask for quotes.¹

The system found the facts and connected them to the buyer’s priorities. That's the result I care about.

3. Proven

A lot of AEO advice gets weak when the subject turns to proof.

A business can say anything about itself.

“We are the best.”

“Oklahoma’s leading experts.”

“World-class service.”

Those statements are cheap. What evidence gives a search or answer system a reason to take the description seriously?

Google’s current generative-search guidance recommends unique, expert-led, non-commodity content instead of pages rewritten only for AI systems.³

Microsoft is even more direct in its AI Performance guidance for Bing Webmaster Tools. It recommends depth and expertise, current information, less ambiguity, and claims supported with examples, data, and cited sources when content may be reused in an AI answer.⁸

Microsoft’s search team frames the underlying problem as a shift from finding the best document to finding information that can responsibly support an answer. It describes the useful unit as groundable information: specific, supportable facts with clear provenance.⁹

It lines up with what I've seen.

A page saying “we are great at SEO” gives the system almost nothing. A page showing the search query, Search Console data, ranking progression, date, URL, and recorded result gives it evidence.

A contractor saying “we build quality showers” is one thing. Completed projects, waterproofing details, specific construction methods, experience, reviews, and project photographs are something else.

Reviews matter too. Google says review quantity and positive ratings can contribute to local prominence. Links and other signals of how well known a business is also influence prominence.⁷

I wouldn't turn that into the claim that reviews are a disclosed AI Mode ranking factor. They aren't. Google has confirmed that reviews help its understanding of local prominence, and it has confirmed that its AI search experiences build on existing Search systems. That's enough reason to strengthen legitimate review acquisition as part of the same foundation.

Claims describe a company. Proof gives the description weight.

4. Shortlisted

Now we get to the part a business owner actually cares about.

A site can be crawled and indexed. The system can understand exactly what the business does. The site can even be cited. None of that guarantees the company will appear when a buyer asks who to call.

Making that answer is the short list.

I've recorded different levels of shortlist behavior across my own work.

In one unbranded ChatGPT test, I asked, “Who’s doing real local marketing work in Lawton, OK?” I didn't name 580 Digital Infrastructure.²

ChatGPT discussed several local agencies, then made a separate section called “One I would watch closely.” The section was about 580 Digital Infrastructure. It correctly identified the combination of local search, websites, direct mail, local campaigns, lead capture, and follow-up systems. It also referenced visible campaign work and a specific SEO case study.

The answer eventually placed 580 Digital Infrastructure among the businesses it considered at the top of the conversation for serious local-growth work in Lawton.²

I ran another test in Google AI Mode using the query “Local marketing Lawton OK.” Google independently surfaced 580 Digital Infrastructure and described the company around print campaigns, web development, Google visibility, and lead-follow-up systems.¹⁰

I followed up by asking which option it would call first for a local small business in Lawton. In that recorded conversation, Google AI Mode placed 580 Digital Infrastructure first.¹⁰

Google hasn't officially ranked 580 Digital Infrastructure as the best agency in Lawton. In that recorded AI Mode interaction, after considering the businesses it found, 580 Digital Infrastructure made the short list and was placed first for that particular buyer profile. That's the whole claim.

The evidence is already good. Exaggerating it would only make it weaker.

A citation isn't a recommendation

These terms get mixed together constantly, and they should not.

Scentonym, a property I built, recorded 35,310 AI citations across 181 consecutive days in Bing Webmaster Tools, with an average of 59.75 cited pages per day.¹¹

That's a large body of measured AI visibility. It doesn't mean 35,310 customers visited the website. It doesn't mean the site was recommended 35,310 times. It doesn't tell us where a citation appeared in each answer.

Microsoft defines Total Citations as the number of times content is displayed as a source in supported AI-generated answers. Microsoft says the metric doesn't indicate placement in the answer. Average Cited Pages doesn't indicate ranking or authority either.⁸

There are several different outcomes worth measuring:

Outcome	What it tells us
Found	The system retrieved or surfaced the business or content
Cited	The site appeared as a visible source
Matched	The system connected the business to the user’s need
Recommended	The business was presented as an option
Preferred	The business received unusually strong consideration or priority

Research is beginning to draw similar lines. A 2026 preprint analyzing 21,143 valid search-layer citations separated citation selection from citation absorption. The researchers distinguished between a page being selected as a source and its evidence actually shaping the final answer.¹²

I find that much more useful than counting mentions and declaring victory.

A business can clear three steps and still disappear

The most useful thing I've learned from testing local AEO is how easily a business can disappear.

One question can produce an unusually strong recommendation for 580 Digital Infrastructure. Another question that feels almost equivalent can leave it out completely. Narrow the need slightly and the company appears again.

It doesn't automatically mean something is broken. It may show how much the retrieval path changes with the wording of the question.

Different phrasing can change the searches being run, the sources being retrieved, and the companies that make the final answer. Current research on paraphrase sensitivity supports treating AI visibility as a probability to measure, not a permanent ranking to claim.⁵

A broader 2026 survey of 45 GEO and AEO studies reached a cautious conclusion. The field still lacks evidence for one universal technique that creates stable, long-term, cross-platform organic visibility and business results. The author recommends repeated measurements, paraphrases, and controls instead of treating one screenshot as a permanent position.¹³

I agree with them.

My goal isn't to make ChatGPT say a company is number one once. I want to increase how often the company enters the right short lists across the kinds of questions real customers are likely to ask.

We can measure that.

The companion paper, Measuring AEO, defines the observation rules, recommendation profile, and intent-normalized metric I would use to do it without turning one favorable answer into a permanent ranking claim.

What appears to help

I don't think anybody outside the companies building these systems knows the complete formula. Anybody who claims to know it should be able to prove it.

The evidence is clearer around a few fundamentals.

The business needs to be technically accessible and indexable. Google and OpenAI both say crawl access matters for inclusion in their respective search experiences.³⁴

The available information needs to match the customer’s actual need. A controlled 2026 study involving 252,000 comparisons across six language models found topical relevance was one of the strongest factors in which retrieved source received the first citation. Completeness and trust cues helped more modestly. Formatting changes by themselves produced little consistent impact.¹⁴

The business needs facts instead of interchangeable adjectives. Its local information needs to be accurate. Legitimate reviews help Google understand prominence. Original expertise and firsthand evidence give search systems information they can't pull from any generic page.³⁷⁹

The rest of the web should tell the same story about who the business is.

I look at the website, Business Profile, reviews, citations, case studies, and third-party mentions as one connected system. They aren't isolated marketing chores.

AEO should not make your content worse

Google has already pushed back on several popular AEO tricks.

Google says you don't need llms.txt for Google Search. You don't need special AI markup. You don't need to break every page into tiny “AI-friendly” chunks. There is no special AEO schema required for Google AI features.³

Google has also warned against manufacturing pages for every possible fan-out query just to manipulate visibility in generative search.³

Good. The internet doesn't need more pages written for robots by robots.

The better approach takes more work. Know something. Explain it clearly. Show the evidence. Make the business easy to identify. Keep the information accurate. Build enough of a reputation that the claims don't stand alone.

A machine can do something with that. So can a person trying to decide who to call.

How I measure local AEO

I wouldn't sell a business an AEO campaign, disappear for six months, and come back with one flattering screenshot.

The baseline comes first.

Start with questions real customers might ask. Do not use twenty small variations of the same keyword. Use actual buying situations.

Record which businesses appear. Separate simple mentions from active recommendations. Note what the system believes each company does and which sources it uses. Test several phrasings of the same intent, then repeat the test later.

Make the improvements and measure again.

The questions should fit the business. A roofer needs tests around storm damage, insurance work, roof replacement, metal roofing, emergency repairs, and contractor reputation. A restaurant belongs in a completely different set of decisions.

For a marketing agency, I want to know whether the system understands the difference between somebody selling social-media posts, somebody selling radio inventory, somebody building websites, and somebody building an acquisition system that connects several channels.

AEO measurement should follow customer decisions, not vanity prompts.

So what is Answer Engine Optimization?

Here is the definition I use:

Answer Engine Optimization is the work of improving the odds that a business is found, correctly matched to a buyer’s need, supported by credible evidence, and included on the short list in AI-assisted search.

For Google, much of that work is still good SEO. Google says so itself. The output is what changed.³

Traditional search asks where a page ranked.

AI visibility adds more questions:

• Did the system find the business?
• Did it understand the business correctly?
• What evidence did it use?
• Did the business make the short list?

Those are marketing questions now. They will matter more as customers become comfortable asking AI systems who they should hire and where they should spend their money.

The standard I use at 580 Digital

I don't believe anybody can honestly guarantee that ChatGPT, Google AI Mode, or another answer engine will recommend a particular business. OpenAI says there is no way to guarantee top placement in ChatGPT Search, and the current research does not establish a universal method for stable cross-platform visibility.⁴¹³

What I can do is document the starting point, improve the search and information foundation around the business, make its expertise easier to understand and verify, and measure what changes.

Most importantly, I can show the work.

I have properties earning top organic search positions. I have an information property generating tens of thousands of measured AI citations. I have local businesses appearing in unbranded AI buyer questions. I also have examples where the system does more than mention the business. It explains why the company belongs on the short list.¹⁵¹¹¹²

That's the standard I want to keep building against.

The question isn't whether we can make an AI repeat the company name.

The question is whether the business makes the short list when the question matters.

That's worth optimizing for.

References and evidence

Official platform documentation

3Google Search Central. “Google’s Guide to Optimizing for Generative AI Features on Google Search.” Google. Accessed August 9, 2026. Primary support for Google’s descriptions of RAG, query fan-out, Search eligibility, the relationship between SEO and AEO/GEO, Business Profile accuracy, unique content, llms.txt, content chunking, special markup, and scaled-content abuse.

4OpenAI. “ChatGPT Search.” Updated July 30, 2026; accessed August 9, 2026. Primary support for ChatGPT Search query rewriting, targeted searches, location use, OAI-SearchBot access, citations, and the statement that top placement cannot be guaranteed.

7Google Business Profile Help. “Tips to improve your local ranking on Google.” Google. Accessed August 9, 2026. Primary support for relevance, distance, prominence, complete business information, links, review quantity, and positive ratings in local ranking.

8Madhavan, Krishna; Meenaz Merchant; Fabrice Canel; and Saral Nigam. “Introducing AI Performance in Bing Webmaster Tools Public Preview.” Microsoft Bing Blogs, February 10, 2026. Primary support for the definitions and limitations of Total Citations, Average Cited Pages, grounding queries, and page-level citation activity, plus Microsoft’s guidance on expertise, clarity, evidence, freshness, and local information.

9Madhavan, Krishna; Knut Risvik; and Meenaz Merchant. “Evolving Role of the Index: From Ranking Pages to Supporting Answers.” Microsoft Bing Blogs, May 6, 2026. Primary support for the distinction between retrieving documents for search and retrieving groundable information that can support an AI-generated answer.

Google Search Central. “AI Features and Your Website.” Google. Accessed August 9, 2026. Additional official guidance on AI Overviews, AI Mode, query fan-out, Search eligibility, internal links, Business Profile accuracy, and the absence of special technical requirements. The newer generative AI optimization guide is the primary Google source used in the report.

Madhavan, Krishna. “Optimizing Your Content for Inclusion in AI Search Answers.” Microsoft Advertising, October 8, 2025. Additional official Microsoft guidance on crawlability, metadata, internal links, semantic clarity, evidence, headings, lists, tables, and keeping important information in HTML rather than only in images or PDFs.

Published and preprint research

5Jack, Will; Noah Lehman; Keller Maloney; and Sarah Xu. “Paraphrase Brittleness in Production Retrieval-Augmented Commercial Recommendation: Reproducibility Below the Rerun-Stability Baseline.” arXiv:2605.27440, May 2026. Preprint. Reports approximately 6,000 paraphrase runs and 6,000 same-prompt controls, with 14 to 29 percent recommendation-set overlap across paraphrases versus a 50 to 61 percent same-prompt baseline.

6Żatuchin, Dmitrij. “Who Owns the AI Recommendation? A Multi-Industry Empirical Map of Brand Category Ownership Across Large Language Models.” arXiv:2606.23057, June 2026. Preprint under review. Examines 3,750 responses across three models and reports 41.6 percent cross-model agreement on the top-recommended brand in its sample.

12Zhang Kai; He Xinyue; and Yao Jingang. “From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms.” arXiv:2604.25707, April 2026. Preprint. Analyzes 602 prompts, 21,143 valid search-layer citations, and 23,745 citation-level feature records while distinguishing source selection from answer-level absorption.

14Vishwakarma, Rahul; Shushant Kumar; and Ratnesh Jamidar. “What Gets Cited: Competitive GEO in AI Answer Engines.” Proceedings of SIGIR 2026, DOI 10.1145/3805712.3808445. Reports 252,000 controlled trials across six language models and 18 content factors. Its controlled test begins after two candidate documents have already been supplied, so its results address citation selection within that context, not organic discoverability.

13Martinez, Olivier. “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026).” arXiv:2607.14035, July 2026. Preprint. Reviews 45 studies and finds no reviewed technique with a demonstrated stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior.

580 Digital first-party field records

1Killian, Jesse. “Parsons Stone & Tile Consumer Recommendation Test.” 580 Digital Infrastructure field test, recorded August 9, 2026. Unbranded ChatGPT query: “Who has the best custom showers in OKC?” Original capture: Parsons Stone and Tile AEO 2.pdf. The public proof page provides the complete readable response and the original source file.

2Killian, Jesse. “580 Digital Infrastructure Local Marketing Recommendation Test.” 580 Digital Infrastructure field test, recorded August 9, 2026. Unbranded ChatGPT query: “Whos doing real local marketing work in Lawton Ok?” Original capture: 580 Digital Infrastructure Local Marketing AEO.pdf.

11Killian, Jesse. “Beyond Citation Count: What 35,310 Reported AI Citations Reveal About Generative Search Visibility.” 580 Digital Infrastructure Research, Research Report 003, Version 1.0, August 2026. The complete overview export spans February 11 through August 10, 2026. An earlier August 9 dashboard capture displays the rounded 34.5K and 59 interface values.

10Killian, Jesse. “580 Digital Infrastructure Google AI Mode Call-Order Test.” 580 Digital Infrastructure field test, recorded August 2026. Initial query: “Local marketing Lawton OK.” Follow-up: “IF you were looking at those options what would be your call order for a local small business in Lawton?” Google AI Mode placed 580 Digital Infrastructure first in that recorded interaction. The public proof page provides the complete stitched sequence from the original captures.

15Killian, Jesse. “Parsons Stone & Tile Search Visibility Case Study.” 580 Digital Infrastructure. Supporting first-party record for the site, query, capture dates, and observed search visibility.

Additional background source

Chen, Mahe; Xiaoxuan Wang; Kaiwen Chen; and Nick Koudas. “Generative Engine Optimization: How to Dominate AI Search.” arXiv:2509.08919, September 2025. Preprint. Useful background on engine differences, prompt sensitivity, and source types. It is retained in the working bibliography but is not currently the sole support for any claim in this report.