Why Did AI Stop Recommending Your Company?

A Drop Out of AI Answers Is a Reading, Not a Verdict

A lost AI recommendation is a sustained fall in how often an AI engine names a company for a buyer’s question, measured across repeated runs. A single check cannot show one. The same question asked twice can return different companies, because every major engine builds its answers fresh each time. Digital Strategy Force treats the first response to “ChatGPT stopped recommending us” as a measurement, not a rewrite.

The stakes explain the panic. McKinsey’s AI discovery research found 50 percent of consumers already use AI-powered search. On the business side, Gartner’s May 2026 survey of B2B buyers found 45 percent used generative AI during a recent purchase, mostly to gather information on vendors and products. When the engine stops naming a company, it is often dropping out of the first conversation a buyer has about the category.

The same McKinsey research explains why a drop can happen without anyone at the company touching anything. A brand’s own sites are only 5 to 10 percent of the sources AI search references; the rest are other people’s pages. An answer is assembled mostly from evidence the company does not control, so the evidence can move while the website stands still.

A typical report reaches the leadership team as a screenshot. A sales director asked ChatGPT for the best vendors in the category, saw three competitors, then saw no mention of the company. The screenshot is real. What it cannot show is how often that same question names the company on any other run, which is the only number that decides whether anything was lost at all.

Every reported drop gets the same treatment. First establish whether the drop is real. Then establish where it happened. Only then spend money on fixing it, because the four places a recommendation can be lost call for four different repairs, and a repair aimed at the wrong place changes nothing.

One Check Is One Sample From a Moving System

Every major engine says, in its own documentation, that its answers vary. Google’s guidance on AI features notes that AI Mode and AI Overviews may use different models or techniques, so the links they show will vary. OpenAI’s developer documentation states that model outputs may differ from request to request by default. Microsoft describes AI citation patterns as inherently dynamic, shifting with models, freshness and changes across the web.

Research measures what that means for a brand. A 2026 study of recommendation behaviour by Malthouse and colleagues found that repeated responses to the same query can produce different brands in a different order. Another 2026 paper ran identical queries five times, then asked how much of that five-call set a single call recovers; the median single call recovered 80.0 to 92.5 percent of it in February, rising to 90.0 to 100 percent in September, across 3,750 records per wave.

Read the other way, a single check routinely misses a brand that a fuller sample would have found. An executive who asks an assistant once, sees a competitor named instead, then calls an emergency meeting has measured one draw from a distribution. The draw may be unlucky. The distribution may not have moved at all.

The variation has two ordinary sources. The engines sample their wording rather than retrieving a fixed reply, so two runs can phrase the same answer around different examples. They also retrieve live sources that change from week to week, so the evidence behind an answer is never quite the same twice. Neither is a malfunction that can be appealed; both can only be measured.

This is the core difference from a Google traffic drop. Search Console reports an average position for every query, so a fall in clicks is usually a real event worth triaging the way the Google traffic diagnosis lays out. An AI answer has no stable position to report, which means the baseline has to be built before a drop can even be defined.

Prove the Drop Before You Chase It

Proving a drop takes a baseline: a fixed set of the questions buyers actually ask, each run several times on each engine across several days, recorded as a mention rate rather than a yes or no. The spread of those runs is the normal band. A later reading counts as a drop only when it falls below that band and stays there on the next round.

A worked example makes the rule concrete. Suppose a company is named in 14 of 20 runs of one buyer question on ChatGPT, with daily readings between 12 and 16 of 20 across a week. A later week at 13 of 20 sits inside that band; nothing happened. A week at 6 of 20, repeated the following week, is a loss worth placing on the matrix. These figures are illustrative, yet the method is exactly what a real baseline does.

Small movements inside the band are not evidence of anything. A 2026 analysis of citation measurement by Sielinski showed that a domain moving from 8 to 11 percent of SearchGPT citations cannot be credited to anything it did, because a 3-point move falls within the typical confidence interval for such a domain. The same logic applies in reverse: a 3-point dip proves no loss.

The baseline also needs to cover more than one engine. ChatGPT, Gemini, Google’s AI Overviews, AI Mode and Perplexity each use their own crawler, index or model. Running the same prompts on each turns a vague worry into two concrete readings, which together point at where the loss happened.

A single check is a sample. A drop is only real when it leaves the band.— Digital Strategy Force, Search Intelligence Division

Most companies do not have this baseline when the first drop is noticed, so the first week of any response is spent building it. That is not wasted time. Every later decision, from whether to act to which layer to fix, depends on knowing what normal looked like.

The DSF Lost Recommendation Matrix

The DSF Lost Recommendation Matrix places a drop out of AI answers by two readings taken from the baseline, so that each combination names the layer that moved. The first reading is shape: whether the mention rate is flickering inside its band, stepped out of it on one date, or sliding down over weeks. The second is scope: whether one engine dropped while the others held, or every engine dropped together.

Shape separates noise from loss and fast causes from slow ones. A flicker inside the band is noise, so the answer is to keep sampling, not to change the website. A sudden step points to something that changed on a date: a model release, a new answer format, a crawler rule. A slow slide points to evidence that is ageing, or to sources that are being replaced by newer ones.

Scope separates the engine’s side from the company’s side. When one engine drops while the others hold steady, the cause lives inside that engine. When every engine drops together, the cause is almost always on the company’s side, because the engines share very little except the website and the third-party evidence they all read.

Take a manufacturer whose name disappears from ChatGPT on one date while Gemini plus Google’s AI answers keep naming it. The shape is a step; the scope is one engine. The first test is that engine’s release notes for the date, then its crawler activity in the server logs, not a content audit. Had every engine dropped on the same date instead, the first call would go to whoever last changed the firewall.

Read together, the two readings narrow a vague crisis to one of six cells, and each cell has a first test that costs hours rather than months. The rest of this guide walks the layers the cells point to, in the order a diagnosis usually meets them: access, retrieval, memory, then format.

Access: The Engine Can No Longer Reach You

The fastest way to vanish from an AI answer is to stop the engine reading the site. OpenAI’s crawler documentation is explicit that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. Perplexity’s guidance likewise asks sites to allow PerplexityBot in robots.txt to appear in its results.

These blocks are often set by accident. Since July 2025, Cloudflare asks every new domain whether to allow AI crawlers, after more than one million customers had already chosen its one-click block. A migration to a new host, a security rule copied from another property, or a CDN setting nobody reviewed can switch an engine off overnight, which is exactly the sudden step this layer produces.

The check for this layer is short. Fetch robots.txt the way each crawler would, list every CDN or firewall rule that touches bots with the date it was set, then confirm in the logs that each crawler still receives normal responses. Most access failures surface within minutes of that sequence, long before anyone needs to discuss content.

Not every rule that sounds relevant is. Google states that its Google-Extended token does not affect a site’s inclusion in Google Search, nor is it used as a ranking signal there. The test that matters for Google is basic eligibility: a page must be indexed and eligible to show with a snippet to appear as a supporting link.

The repair is equally direct. Read the robots.txt file, the CDN and firewall rules, plus the date of the last migration, against the date the step appeared. Server logs settle the question: if a crawler stopped arriving, the crawler log audit will show the day it stopped.

Retrieval: The Pages It Reads Moved On

When the engine can still reach the site but the mention rate slides, the evidence it retrieves has usually changed. A 2025 study from Ruhr University Bochum with the Max Planck Institute for Software Systems ran the same queries two months apart. Only 18 percent of the pages cited by AI Overviews were common to both runs, against 45 percent for organic results.

That churn is the slow slide in its natural habitat. The sources an engine leans on for a category are replaced by newer reviews, newer comparisons, newer coverage. A company that was named because a handful of independent pages described it well can fade from the answer as those pages age out of the set, with no change to its own site at all.

Consider a professional services firm named for a year because two independent comparison guides listed it among the leading providers. When newer guides replace both, without the firm in them, its mention rate slides over several weeks while its own site is untouched. Rewriting its service pages would do nothing. Earning a place in the guides the engine now reads is the repair.

This is why the repair for retrieval is rarely more pages on the company’s own domain. With own sites at only 5 to 10 percent of the sources AI search draws on, the lever is the third-party evidence: current comparisons that include the company, independent reviews, data others cite. The difference between memory and live retrieval decides how quickly any of it registers.

The diagnostic test is a citation comparison. Record which sources the engine cites for the lost question now, then compare them with the sources it cited when the company was being named. The gap between the two lists is the work.

Memory and Format: The Engine Itself Changed

Sometimes nothing changed on the company’s side and nothing changed in the sources; the engine itself did. On February 13, 2026, ChatGPT retired GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini and GPT-5. On May 19, 2026, Google made Gemini 3.5 Flash the default model for the Gemini app and for AI Mode in Search. A new model learned its picture of every category from new data.

A model change produces the classic single-engine step: one assistant stops naming the company on a date that matches a release, while the others hold. Memory is the hardest layer to repair quickly, because nothing a company publishes today enters a model already trained. The detail of how long that lag runs belongs to the refresh clocks behind AI answers; the practical point here is to stop rewriting the website in response.

The answer can also change shape. On shopping-intent questions, ChatGPT now shows product options with images, details and links. In October 2026 OpenAI said ChatGPT ads were running in 89 countries for more than 100,000 advertisers, while Google reported that sponsored and organic product listings in AI Mode were drawing similar click-through rates.

A company can fall out of a format change without losing anything the matrix would call a recommendation. If the answer now leads with product cards or sponsored listings, the paragraph of named vendors may shrink or disappear. The test is to read the answer itself, not just to search it for the company’s name.

Memory and format losses look alike on a chart, since both arrive as a step on one engine, yet they read differently on the page. After a memory change the answer still names vendors, only different ones. After a format change the answer names fewer vendors in its text, because cards or sponsored units now carry the commercial part. Logging the answer text from each baseline run, not just the yes or no, keeps the two apart.

Repair the Layer That Moved, Then Measure

Each cell of the matrix has its own repair and its own clock. An access fix shows within days, once the crawler returns. A retrieval fix takes weeks, because new evidence has to be published, indexed and chosen. A memory change can outlast every fix until the next model, which is why independent evidence that retrieval can find matters more than anything written for a model’s memory.

The measurement no longer has to be improvised. Since June 2026, Search Console reports impressions inside AI Overviews and AI Mode. Bing Webmaster Tools reports how often each URL is cited in generative answers, and since June it also shows citation share against compared sites. Google’s own figures on how often a site appears in AI answers make the baseline easier to anchor.

First-party reports still do not replace the repeated-prompt baseline, because they show impressions and citations rather than whether the company is the one named for the buyer questions that matter. The two together make a drop readable within a day of it starting. They also protect a company from the opposite mistake: treating ordinary flicker as a crisis, then renting visibility it does not need.

Digital Strategy Force runs this as a standing programme for clients rather than an emergency service, because a baseline built before the drop is worth far more than one built after it. The order never changes: prove the drop, place it on the matrix, repair the layer that moved, then measure the next run against the same band.

The same band that proved the drop also judges the fix. A repair that returns the mention rate to its old band is working. One that lifts it a few points while it is still below the band has proved nothing yet, for the same reason a few points of loss proved nothing at the start. Patience here is not caution for its own sake; it is how a company avoids paying twice for one problem.

FAQ — Why AI Stopped Recommending Your Company

Next Steps — Build the Baseline Before the Next Drop

▶ Write down ten buyer questions your company should be named for, and run each one at least five times on every major engine to set the baseline band.

▶ Re-run the set on a fixed schedule, logging the date of every model or product change the engines announce.

▶ When a drop leaves the band, read its shape and scope on the DSF Lost Recommendation Matrix before touching the website.

▶ Check access first if every engine dropped on one date: robots.txt, CDN and firewall rules, plus the last migration.

▶ If the slide is slow, audit which third-party sources the engines now cite, then fund independent evidence where your company is missing.

Building the baseline takes a week; reading a drop against it takes an afternoon. Keeping that baseline running across ChatGPT, Gemini and Google’s AI answers is standing work in Answer Engine Optimization. It starts with one question nobody has measured yet: how often is your company named for the questions your buyers ask, and what does normal look like?