Six chapters, the method behind every number, and a plan for each AI model. Enter your work email to open it.
By continuing you agree to our Privacy Policy.
Peek · AI Discovery Black Book · Third Edition
A renter tells AI: two bedrooms, under $2,200, near work.
ChatGPT runs about 11 searches they never see.
The AI Discovery Black Book shows what they find.
We spent nine months watching ChatGPT, Claude, Gemini and Perplexity recommend apartments. Here's what we learned.
Up every month since May, to 73% in August, the highest in our three editions. Nearly 9 in 10 trace back to Apartments.com and Zillow.
ILS share of all AI citations, Jan–Aug 2026. December is left out because a model change skews it.
How to cover the right listing sitesZillow's network includes Redfin and Trulia. CoStar's includes Apartments.com and Homes.com.
Share of listing-site citations, Jun–Aug 2026. Yardi's RentCafe (8%) runs on its own database.
Which listings to prioritizeCoStar's network closed to within 6 points this spring. This summer Zillow pulled away again, and Redfin, which carries Zillow's rentals, is now cited as often as zillow.com.
Share of listing-site citations by network, per month, Dec 2025 – Aug 2026.
What it means for your listingsOver nine months, when a property's AI visibility jumped and stayed up, the new citations came from listing sites, not its website.
Lasting gain: up 7+ points and still up a month later. Example: a 150–249-unit community in Louisville, CO, June → July 2026.
What drove the gainsTesting against a single prompt is incomplete at best. LLMs take one question and break it up, and you need to measure coverage against each one.
ChatGPT's average background searches per question. Gemini runs 6, Claude 2. Every four-model assessment, August 2026.
How each model searchesThe most of any model. Claude used to lead, but from June to August Claude slid from 26% to 22% while ChatGPT rose from 22% to 28%.
Property-website share of each model's citations, per month, Jun–Aug 2026.
A plan for each AI modelProperties with clear rent and commute info get recommended more when renters ask about them. That advantage has shrunk since winter, from 45% to 20%. But 64% of properties still fall short on both, so it's still an easy win.
How much more often the five properties AI recommends most in each market show up than everyone else.
What to put on your pagesAsk AI for an apartment and 80% of answers name a specific community. Only 14% name a management company. Corporate homepages get under 0.3% of citations.
Everyday apartment questions in our test that don't name a company. Each dot is 1% of answers.
Which owners are winningThe major platforms land within a few points of each other. Switching vendors won't change how often AI recommends you.
Market Performance (0–100): visibility vs. the top property in the same ZIP, Dec–Aug, platforms with 30+ properties. A single property swings about 8.5 points on its own.
What to put on your site“Reddit drives AI answers.”
“AI checks your reviews first.”
“My website is blocking AI.”
Searches: ChatGPT, Gemini and Claude, August 2026. Crawler check: property websites, Jun 10 – Aug 31. Each dot is one site we checked.
Up from 25% in February. More answers point to one layout ("the two-bed B2") or one unit ("#412"), not just a building. An early signal we'll keep tracking.
Out of about 4,000 ChatGPT answers a month.
What to watch nextSix chapters, the method behind each number, and a plan for each AI model.
Zillow's lead is back. Claude is no longer the easiest model to break into. And the content that moves a property most hasn't changed, but its edge is half what it was.
Throughout, we measure visibility with AI Discovery Score: the share of relevant AI answers where a property shows up. Think organic search ranking, but for AI.
searches behind one ChatGPT question
You're marketing to two audiences now. The renter asks the question, with a budget, a commute and a must-have list. Their preferred AI model then runs its own searches in the background (about 11 for ChatGPT, 6 for Gemini, 2 for Claude) and favors the properties whose pages answer the most of them.
See one tracedof AI answers recommend a specific community, not a brand
Renters don't ask AI for a brand, and AI doesn't answer with one. On everyday apartment questions in our test, 80% of answers recommended a specific community. Only 14% named a management company.
Owners and brandsZillow's lead over CoStar this summer, up from 6 in spring
Apartments.com and its sister sites were closing the gap on Zillow's lead in AI citations this spring, but that trend has reversed. Zillow's network climbed back to 54% of ILS citations this summer, 21 points ahead. More of that Zillow data now reaches AI through Redfin, which is now cited as often as zillow.com.
The listing storyof ChatGPT's citations are property websites, the most of any model
Perplexity (81%) and Gemini (73%) still lean hardest on listing sites. Claude used to be where a property's own website had the best shot, but its ILS share climbed from 50% in February to 63% this summer. ChatGPT has taken its place: property websites went from 22% of its citations in June to 28% in August, while Claude slid from 26% to 22%. Perplexity was at 12% and Gemini at 6%.
The four modelsof properties fall short on rent and commute info
Properties with clear rent and commute info get recommended more when renters ask about them. That edge has halved since winter: 20% more often on price questions (down from 45%) and 18% on commute (down from 35%). Most properties still haven't caught up, so there's room to gain.
Check your propertyis all that separates the major website platforms
Every major website platform lands within 10 points of the others on Market Performance, our score for how visible a property is compared with the best property in its market. Switching vendors won't change how often AI recommends you.
Platforms comparedcarried six of seven lasting gains
We tracked the same properties for nine months. When a property's score jumped and held, the added citations came mostly from listing sites like Zillow, Trulia, Redfin and Apartments.com, not its own website: six of the seven lasting gains. Drops were split: of the two that held, one ran through listing sites and one through the property's own website. Keep every listing complete on both networks.
One property, traced"Reddit drives AI answers."
of AI citations come from Reddit, every month since December.
"AI checks your reviews first."
of AI searches in August looked up reviews. The models searched price and commute.
"My website is blocking AI."
property sites blocks an AI crawler. Almost none do.
Searches: ChatGPT, Gemini and Claude, August 2026 (Perplexity doesn't show its searches). Crawler check: property websites, Jun 10 – Aug 31.
of what AI cites for apartments comes from listing sites. The highest of our three editions.
Nearly 9 in 10 of those citations trace back to two databases.
Listing sites, which the industry calls ILSs, are Zillow, Apartments.com and the like. When AI answers a renter, it shows links to its sources; each link is a citation. Most of those citations point at listing sites. Being on Zillow, Trulia and Redfin looks like three listings. It's one source: all three pull from Zillow's database. Syndication deals mean most listing sites are really windows onto one of two feeds.
Share of listing-site citations, Jun–Aug. Yardi's RentCafe (8%) runs on its own database.
Scroll through the year.
Share of listing-site citations by network, per month.
Zillow's network supplied 57% of listing citations. CoStar had 28%.
Zillow dropped to 45% while CoStar rose to 39%, the closest the two have been.
In June we said the summer would tell us. It did: Zillow back to 54% against CoStar's 32%, a 21-point lead.
Redfin carries Zillow's rental listings, and AI now cites it as often as zillow.com itself. From 2027, after the FTC settlement, Redfin can sell rental advertising on its own.
You don't need a listing on every site. Zillow's network includes Redfin and Trulia. CoStar's includes Apartments.com and Homes.com. Listings on both networks cover 86% of the listing data AI reads. Add Yardi's RentCafe, which runs on its own database, and you've covered almost all of it.
We tracked the same properties for nine months. On identical questions a score moves about a point in a typical month and 4–5 points one month in ten, so we only counted moves of 7+ points that held the next month. Here is one.
Its score jumped. Its listings carried more of the citations. Its own website carried fewer.
When a property's AI visibility jumped and stayed up, the listings were behind it six times out of seven. Your website still matters, especially on ChatGPT, which cites property websites more than any other model.
searches run behind a single renter message on ChatGPT. The renter never sees them.
Each model searches, and cites, differently. One plan won't cover all four.
The model folds what it knows about the renter into a few short searches, reads the results, and recommends from what those pages say. Here is a real one, traced.
Search strings from a real ChatGPT response in the August sample, shortened. Community A and B are illustrative.
A budget, a commute, two bedrooms, usually spread across several back-and-forth messages. Here is one of them.
In the background. It averages about 11 searches per question. Gemini runs 6, Claude 2.
The budget shows up in 44–59% of searches behind a price question; the commute in 61–79% of responses to a commute question. Reviews and Reddit: no model searched for them in the August sample.
The page that states a rent and a drive time is the page the model can cite. A page without them drops out.
The citations come from those searches: 99.7–100% of the sites ChatGPT and Claude cite showed up in their own results.
Long messages, short searches. In a small test, we wrote renter questions as full paragraphs, about 84 words each: the dog, the commute, the budget, family nearby. The searches AI ran from them averaged 8 words. One 96-word message turned into two searches for average rent in Dallas and Grand Prairie. The dog, the office, the budget and the sister's neighborhood all got dropped.
One surprise. In a separate test of questions written the way renters talk, the most-searched word was "2026." None of the property pages checked use it.
August 2026: 295 four-model assessments, 1.4M background searches. The "2026" finding is from a data science test: 71 questions, 144 searches, one model. The long-message test: 6 paragraph questions (77–96 words), 2 runs each, one model with live web search.
Perplexity gets 81% of its citations from listing sites, and Gemini 73%. Claude, once the easiest opening for a property's own website, climbed from 50% in February to 63% this summer. In August, ChatGPT leaned on listing sites the least of the four. Why it matters: the model your renters use changes where you need to show up.
Listing-site share of each model's citations. Comparisons start in February, the first full month after the Jan 15 change to prompts and citation capture. Perplexity starts in June, after a change to how its citations are captured.
Each model pulls from different places, so each needs a different fix. ChatGPT is the one that reads your website, and the only model where properties show up less than last winter.
Put rents and floor plans on your website.
28% of its citations were property websites in August, the most of any model.
Be listed on both the Zillow and CoStar networks.
65% of its citations come from listing sites.
Be listed on both networks, and state drive times.
71% of its citations come from listing sites, and 35% of its searches are drive-time lookups.
Keep every listing complete.
81% of its citations come from listing sites, the most of any model.
of properties still trail the top five on price and commute.
Price and commute still separate the top properties, by half as much as last winter.
Lift is how much more often the top five properties show up than everyone else on a topic. Price and commute still lead, but the gap has narrowed, mostly because Gemini now puts less weight on them: its price lift fell from 68% to 23%.
Amenity questions were regrouped in May, so compare amenity across editions with care.
of everyday AI answers in our test recommend a specific community. 14% recommend a management brand.
Renters don't ask AI for a brand. They describe what they need, and AI answers with a building.
The corporate homepage, the page a brand score tracks, gets under 0.3% of citations. Corporate sites earn citations by hosting community pages. Ownership explains only about a quarter of a property's score; the community explains the rest.
Share of citations, ChatGPT, Claude and Perplexity, Jun–Aug. Gemini's citations carry only the site's domain in this window, not the page. Bars are scaled to the largest value.
Seven new names in the top 12 since June. Portfolio size has nothing to do with who wins: the 10 largest NMHC managers win about 16% of AI's top-five spots while running 13% of the properties AI mentions. Being big earns no bonus. And most of last edition's leaders didn't get worse: the best property in each of their markets got better, which raised the bar.
| # | Owner | Move | Market Performance | Properties |
|---|---|---|---|---|
| 1 | Pantzer Properties | New | 34 | |
| 2 | Monarch Investment and Management Group | New | 240 | |
| 3 | MG Properties | New | 91 | |
| 4 | Continental Properties | ▲ 5 | 100 | |
| 5 | Oakmont Properties | New | 24 | |
| 6 | Olen Properties | New | 53 | |
| 7 | Veris Residential | ▼ 5 | 33 | |
| 8 | Morgan Properties | ▼ 7 | 329 | |
| 9 | AvalonBay Communities | ▼ 4 | 293 | |
| 10 | UDR | ▼ 7 | 175 | |
| 11 | FriedLam Partners | New | 55 | |
| 12 | Sentinel Real Estate | New | 96 |
Move compares rank with our last edition; "New" means not in last edition's top 12. Dropped out since June: Nuveen Real Estate (#13), LivCor (#15), Equity Residential (#19), Berkshire Residential Investments (#22), Mill Creek Residential (#25); The Connor Group and GID had too few properties scored to rank. Properties = Yardi-tracked portfolio.
UDR's Residences at the Domain in Austin showed up about as often this summer as last winter (26% → 27% of answers). But Austin's most visible property jumped from 26% to 68%. So its Market Performance fell from 100 to 40, without the property doing anything worse.
Called MAAR (Market-Adjusted Appearance Rate) in our last edition.
How visible a property is compared with the most visible property in its own market. The leader scores 100. It lets a property in Austin be compared fairly with one in Atlanta.
The top 12 run anywhere from 24 to 329 properties.
Inside a typical portfolio, the best property scores 10 times higher than the worst.
About 7 in 10 of this summer's top-five properties were there the month before.
We fetched 316 property websites the way a crawler does. Almost none block AI. What sets the weaker sites apart is what's missing once AI gets in.
can't answer "what does a two-bedroom cost" from the structured data on the page
carry no FAQ markup
show nothing without JavaScript. They score lower (7.9% vs. 19.9%), though 19 of the 21 belong to one organization, so treat it as a link, not a cause.
block an AI crawler. Some sites turn away simple requests, but AI still gets in.
Crawled Jun 10 – Aug 31. These are organization-level vendor choices; the data shows what sites lack, not that fixing markup raises a score.
is all that separates the major website platforms on AI visibility.
Every major platform lands within 10 points of the others on Market Performance. What your website does control is the detail listing sites don't carry: the real price of a floor plan, the drive to a nearby employer.
of ChatGPT answers now reach a floor plan or a specific unit. In February it was 25%.
People rent a unit. Answers are moving toward one. An early signal, back next edition.
What to watch: whether the pages AI cites, and the pages renters land on, shift from listing sites and community pages to floor plans and units. If they do, the winning page carries unit-level facts: this unit, this rent, available this date, tour it here.
From the renter's side, mock searches in the ChatGPT app named far more specific units than a few months ago, including units already rented. AI is getting more specific before it gets more current.
| Model | Names the property | Communities per answer | Same communities on a repeat ask |
|---|---|---|---|
| ChatGPT | 26% → 36% | 5.7 → 10.6 | 16% → 32% |
| Claude | 24% → 30% | 4.6 → 7.7 | 34% → 41% |
| Gemini | 34% → 35% | 14.3 → 6.6 | 33% → 34% |
Current model → its newer version: gpt-5-mini → gpt-5.6-luna, claude-haiku-4-5 → claude-sonnet-5, gemini-2.5-flash → gemini-3.6-flash. 4,050 answers across 25 properties, each question asked three times, August 2026. In August, the assessments in this report ran the current model in each pair.
What to watch: ChatGPT's newer model names the property more often and names nearly twice as many communities per answer. Gemini's newer model names less than half as many. A score can move when the model changes and nothing about the property does.
Ask today's models the identical question three times, and only about a quarter of the communities they name are the same each time. Each newer model is more consistent. Read one assessment as one draw, and read a trend across several.
Your score on ChatGPT, Claude, Gemini and Perplexity, against the competitors in your market.
Unless noted, everything in this report comes from Peek's AI discovery dataset, December 4, 2025 to August 31, 2026: 1,701 completed assessments of subscriber properties, measured against 8,442 top competitor records. Each assessment runs a set of renter-style prompts across ChatGPT, Claude, Gemini and Perplexity, with web search on, and records which properties get cited and which sources the models use.
Two filters apply to every number. We keep only the top five competitors per assessment (since July 14, Discover records every competitor it sees), and only completed assessments that ran all four models, which drops 129 other scorecards: 53 runs that failed and 76 single-model reports. With both filters off, our scripts reproduce the first two editions' tables.
Windows are named by edition: Ed. 1 · Dec–Feb, Ed. 2 · Mar–May, Ed. 3 · Jun–Aug. Gemini's citations carry only the site's domain in this window, so page-level cuts use ChatGPT, Claude and Perplexity. AI Discovery Score is the share of relevant AI queries where a property shows up.
| Date | Change | Handling |
|---|---|---|
| Jan 6, 2026 | Gemini began naming about six times as many properties per answer | December excluded from per-model comparisons |
| Jan 15, 2026 | Prompt shortened; citation capture moved to each model's gateway | Jan 16 – Feb 28 is the earliest comparable per-model window |
| Apr 10, 2026 | Failed assessments still wrote a scorecard | Four-model filter removes them |
| May 26, 2026 | Perplexity citation extractor rewritten | Perplexity compared Apr vs Ed. 3 only |
| Jul 14, 2026 | Discover records every competitor, not only the top five | Top-five filter restores five per assessment |
Background searches: every four-model assessment in August 2026 (295 assessments, 286,746 responses, 1,405,208 searches). Perplexity doesn't show its searches.
Brand prompts: 200 branded prompts plus 50 unbranded controls on Gemini; a 50-prompt sample on Perplexity.
Website crawl: 316 property websites fetched the way a crawler fetches them, Jun 10 – Aug 31.
Renter-language test: our data science team, 71 questions about Dallas and Austin, one model, 144 searches, checked against 29,897 property pages.
Same-property panel: 264 properties assessed in two or more months, December to August.
Model upgrade test: 25 properties, 9 question types, each asked three times on each model's current and newer version, 4,050 answers, August 2026. Thresholds were fixed before any data was collected.
Peek's recurring report on how AI models recommend apartments. The third edition covers December 2025 to August 2026: 1,701 four-model assessments and 8,442 top competitors across ChatGPT, Claude, Gemini and Perplexity.
Mostly from listing sites. From June to August 2026, 72% of the sources AI cited for apartments were listing sites (ILSs) like Zillow and Apartments.com, the highest share in our three editions. Nearly 9 in 10 of those citations trace back to Zillow's and CoStar's networks.
About 11 on average, in the background, before it answers. Gemini runs about 6 and Claude about 2. The renter's budget and commute often go into those searches; in our August 2026 sample, reviews and Reddit did not.
ChatGPT. From June to August 2026 its share of citations going to property websites rose from 22% to 28%, the most of any model, while Claude's slid from 26% to 22%.
Not measurably. The gaps between the major website platforms are about the size of one property's normal month-to-month swing. What's on the page, like real rents by floor plan and drive times, matters more than who built it.
No. On everyday apartment questions in our test, 80% of AI answers recommended a specific community and 14% recommended a management brand. A corporate homepage gets under 0.3% of AI citations.
Price and commute. On price questions, the top five properties show up 20% more often than the average property, down from 45% last winter, and 64% of properties still trail the top five on both price and commute.
Not in our data. Reddit is under 0.1% of AI citations in every month since December 2025, and ChatGPT, Gemini and Claude ran no searches for reviews in our August 2026 sample.
Earlier editions: First Edition, March 2026 · Second Edition, June 2026.