Plate Lunch Collective

06/10/2026

Getting an AI assistant to understand what your business even is can take real work, and I know this one from the inside. The phrase plate lunch means a food order in the training data, overwhelmingly. Getting AI platforms to read Plate Lunch Collective as an AI search optimization agency in Aiea, Hawaii took deliberate work on the signals that define the entity.

That is the resident's discovery problem in miniature. A Hawaii business name collides with mainland namesakes constantly, and the model often holds a stronger picture of the larger company with the same name, so it answers about them instead of you. There is a second half. A model flattens Hawaii into a single location, but a Maui customer cannot use a Big Island plumber.

The resident economy mostly runs on word of mouth, so a business living on referrals does not feel this until it needs a customer who is not already in the network. The person who just moved here. The one whose usual guy retired. The one who asks ChatGPT instead of asking a friend. For that customer, two things decide whether you surface: whether the tools can tell who you are, and whether they place you on the right island.

This is one of three problems. Visitors plan through AI assistants months out and get handed whichever operator the tools can read cleanly. Off-island buyers search like the national buyers they are and never type Hawaii at all. The resident-facing business has its own, and it comes down to an unambiguous entity on the correct island.

Full read ⬇️

06/10/2026

Apple's new Siri runs on the same two layers as every other answer engine. It ships this fall with iOS 27 as the default answer layer across more than 2.5 billion active Apple devices, the largest installed base in consumer tech.

The parametric layer is what the model already believes about a brand from training, fixed in its weights between runs. If that picture is missing, outdated, or filed under the wrong category, Siri inherits the error before it retrieves a single page, and a freshly published page may not override it. The retrieval layer is what Siri pulls at query time, assembled from passages a model can lift and have stand on their own. The unit that comes back is the passage rather than the whole page.

Apple built its Foundation Models with Gemini technology, confirmed at the keynote. If Siri's web answers are produced by Gemini-derived models, what surfaces is likely governed by retrieval and ranking close to Google's own AI answers rather than a net-new Apple index. The entity signals, structured data, and Knowledge Graph presence that earn visibility in Google's AI answers are already doing part of the work for Siri.

One channel does not run through a website at all. When Siri brokers a local or commercial choice for someone with no prior relationship to the brand, the candidate comes from Apple's own local data. Presence in Apple Maps and Apple Business Connect becomes the lever that decides whether Siri ever says the name to someone new. For hospitality, local service, and tourism, that route is the one that decides whether the name comes up at all.

The open question is the index, whether the Gemini world-knowledge tier retrieves against Google's web index or something Apple supplies. That is the first thing to test when iOS 27 ships.

Full read ⬇️

06/09/2026

$400M+ in venture funding has gone into AI visibility tools, and as of June 6 the tracker lists 56 platforms. With that much capital in the category, three of the hardest measurement problems are still unsolved by every platform on it.

Query fan-out is invisible to all of them. When a model receives a query it decomposes it into sub-queries and retrieves content for each one internally. Whether a brand shows up is decided in those sub-queries, and none of them are exposed to external tools. All 56 platforms measure the primary prompt. The layer where the decision actually happens stays dark.

Most visibility scores are a single draw from a probability distribution. AI outputs are probabilistic, so the same prompt run 100 times returns 100 different responses. SparkToro and Gumshoe, using Carnegie Mellon's LLM-consistency methodology, found less than a 1-in-100 chance that ChatGPT or Google AI repeats the same brand recommendation list across 100 identical runs. A weekly single-run score reports one position inside a wide distribution and presents it as a measurement.

There is no Search Console for AI assistants. (rolling out now to select accounts). OpenAI does not expose query data, so every platform either generates its own prompts or tracks ones you enter by hand. The prompts being monitored are approximations of buyer behavior, not observed queries. Useful, as long as you know that is what they are.

The June update adds Cognizo, Brandi AI, HubSpot's AEO module, and SE Ranking's visibility tracker, and refreshes funding and methodology for the rest. Full breakdown of all 56 in comments.

04/30/2026

We originally published this piece in March 2025 under a different title arguing that retrieval layer optimization is the single strategy that works across all AI interfaces. That argument still holds, but it was incomplete. It addressed one of the two systems that determine what AI says about your brand and ignored the other.

(Full update in 💬)

60% of ChatGPT queries are answered from parametric knowledge without triggering a web search. The model answers from memory. No retrieval runs. No content gets fetched. No citation opportunity exists.

If your brand has no presence in the model's training data, you are invisible on the majority of queries. Not because your content is poorly structured. Because the model never looked.
This revision adds the parametric layer, updates every statistic to current sourcing, and removes claims we could not trace to a primary source.

04/30/2026

Most brands working on AI search optimization are solving the retrieval problem. They structure content well, it gets retrieved, it even gets cited. But the brand is never named in the response.

(Full breakdown in 💬)

Seer Interactive found that 61.7% of all AI search appearances are ghost citations. The URL shows up in the footnotes. The brand is absent from the recommendation. In the worst version, a competitor is explicitly named and recommended in the same response that cites your content as the supporting evidence.

The mechanism matters here. The model is not reading your page, finding it persuasive, and deciding to recommend you. It recommends brands it already knows from parametric memory, then retrieves content after the fact to justify the decision. Your content provides the evidence. Someone else's brand gets the endorsement.

This is why content optimization and entity optimization are two fundamentally different problems. Content changes propagate to retrieval systems within days. Brand mention changes take six to twelve weeks because you're changing what the model's training data knows, not what its search index can find.

If your content is being retrieved and cited but your brand isn't being named, the content is working. The parametric layer doesn't recognize your brand as a credible entity in your category. That requires entity signals: Knowledge Graph presence, consistent naming across authoritative third-party sources, brand-as-subject positioning. Different work, longer timeline, and most teams haven't started it.

04/29/2026

There used to be people you went to. The handyman who'd actually pick up the phone. The urgent care nurse who'd tell you whether the rash needed a doctor. The lawyer friend you'd text when a contractor did something shady. Some people had one or two of these trusted sources. Most people had none. The friction of finding a real expert was, for most of daily life, the friction of having an answer at all.

(Full details in comments - ⬇️)

That friction is gone. All of those people are now one person, and that person lives in your phone.

A couple in Target last month. Mid-aisle, debating something on the shelf. The wife pulled out her phone and said "ChatGPT says" with the same casual authority as "my mom says." Her partner didn't ask which ChatGPT. He knew. Everyone has the same contact now.

Google took about four years to become a verb. ChatGPT did something stranger. It became a contact. Someone you refer to by name and assume the other person knows who you mean.

About 60% of ChatGPT queries get answered from parametric knowledge. The model answers from memory. No URL loads, no referral fires, no analytics event records the influence. Your dashboard was built assuming that decision-influence produces a click. When the influence happens inside a conversation that never touches the web, there is no click. There is nothing to measure. Cloudflare found Claude made ~71,000 page requests for every referral sent back. The platforms consume the web at industrial scale and barely return traffic. What does arrive shows up as "Direct" in GA4 because referrer headers get stripped.

Google searches per US desktop user dropped nearly 20% year over year per Datos/SparkToro. People still use Google. They just run fewer follow-up queries because the AI already handled the synthesis. It doesn't look like people leaving. It looks like people needing less.

Two layers of strategy matter now. Retrieval optimization covers what happens when the model goes looking something up. Parametric presence covers what the model already believes before it goes looking. One without the other leaves you visible in half the picture and invisible in the other half.

04/09/2026

If AI keeps describing your brand the way it used to be, not the way it is now, the instinct is to update your website and wait. That won't fix it.

(More in comments ⬇️)

The misrepresentation is usually happening in the parametric layer. That is the knowledge encoded into the model's training weights, not the retrieval index the model queries in real time. Optimly's analysis of 5,829 brands found that 59.8% of brand misrepresentation errors originate there. This is especially common after a rebrand or category pivot, when the historical footprint still dominates training data.

The way it plays out: the model holds a confident prior from its last training cycle ("this company provides IT staffing"). At query time it retrieves your updated site that now says cybersecurity. It doesn't overwrite the prior. It weighs both. And based on how these systems are architected, the parametric prior often wins because training crawlers process significantly more volume than search crawlers.

Plate Lunch Collective, an AI search optimization agency based in Aiea, Hawaii, starts every engagement with a parametric audit for exactly this reason. There is no point optimizing retrieval signals if the model's trained representation is pulling against you.

The fix requires aligning the authoritative third-party sources that actually feed training data: Wikipedia, Crunchbase, G2, high-authority editorial coverage in your vertical. When those sources consistently describe your brand the same way, the next training cycle ingests coherent signals and the stale weights get corrected. That is entity SEO work. It is the only lever that reaches the parametric layer.

04/08/2026

AI search does not rank your brand. It resolves it as an entity. That is a different process with a different failure mode, and most brands do not know they have failed it until a competitor is being recommended by every AI assistant and they are not.

The resolution process works in three stages. The system extracts the named entity from the query, identifies candidate matches in its knowledge graph, and then weighs surrounding signals to decide which real-world entity the user actually means. That weighing produces a confidence score. If your score falls below the model's threshold, it bypasses your brand entirely during synthesis. It does not rank you lower. It removes you from the consideration set before content quality ever enters the evaluation.

The failure mode that catches brands off guard is that their visibility metrics do not reflect this. If a more prominent entity shares your name, or even operates in an adjacent category with similar language, tracking tools can attribute those mentions to you. Your apparent AI visibility looks healthy. Your actual retrieval presence is zero.

The signal infrastructure required to fix this is specific. An entity home page, typically your About page, that functions as the single anchor where the algorithm establishes your baseline identity. Organization schema with your name, URL, founding date, and stable unique identifiers. Identical information maintained across Wikidata, Crunchbase, LinkedIn, and relevant industry directories. Named co-citations from publications and sources the model treats as authoritative. A founding year that differs by even one year across two platforms is enough for the model to doubt which entity it is resolving. When it doubts, retrieval collapses.

The 2026 ChatGPT entity panel update made the stakes visible in a concrete way. Brands with strong entity signals now appear as clickable interactive panels with summarized facts and trusted links. Brands with weak signals appear as plain text next to those panels. Your buyers can now see that gap directly in the interface when they ask about your category. That is what the entity resolution problem looks like from the outside.

AI search engines resolve brands as entities, not keywords. Learn how entity resolution works and why weak signals erase your brand from results.

04/08/2026

When a CMO or brand manager asks ChatGPT or Perplexity about GLP-1's effect on their category, the AI does not know your brand has exposure. It retrieves from whatever content is indexed, sourced, and specific enough to match the query at the level the question was asked.

That is the structural issue most marketing strategy is not accounting for. AI search does not work like Google's ten blue links. The system breaks a complex question into components and retrieves the best available answer to each component independently. A general statement about GLP-1 and consumer spending does not surface for the specific question about alcohol demand pressure. A section of content that answers that exact question completely, with named research and precise enough language to be attributable, does.

The GLP-1 case illustrates this clearly because the exposure is already documented across 20 verticals. A BMJ cohort study of over 600,000 US veterans found GLP-1 use associated with an 18% lower risk of alcohol use disorder. Numerator household panel data found a 38% increase in protein supplement spending among GLP-1 households. Cornell research documented an 8% decline in fast food spending within six months of GLP-1 initiation. That evidence exists. The brands and research properties that built structured content around it are the ones getting cited when a buyer asks an AI assistant about category exposure.

The compounding problem is the training layer. Large language models are trained on available web content continuously. The positions being established in retrievable content now are shaping what those models answer from parametric memory in 2027. If your category has documented GLP-1 exposure and your content does not reflect it with specificity and attribution, you are not behind on a trend. You are absent from the conversation as the model forms its understanding of the space.

04/06/2026

Updated the BrandLight vs Evertune AEO platform comparison for April 2026.

Both platforms launched activation features in Q1 2026. BrandLight launched AI Ads to track paid placements inside AI responses and Agentic Commerce to monitor how AI agents recommend and purchase products. Evertune launched AI Retargeting via Index Exchange and The Trade Desk to run programmatic campaigns on the publications that feed AI models.

These are not monitoring plays. They are revenue plays. The question for buyers is no longer which tool gives better visibility data. It is which activation layer fits their existing infrastructure.

Gartner published its first-ever Market Guide for Answer Engine Visibility Tools in March 2026. The category has analyst recognition. Enterprise budget is following.

Full comparison at the link