What Happens When AI Has No Local News to Read

Field Notes // Measured August 5, 2026

In June I ran an audit of what the internet says about Johnson City. Three of the five AI systems I tested handed me the same statistic, a rape rate of about 75.6 per 100,000, roughly double the national figure. Three tools agreeing feels like three confirmations. It was closer to three copies of one rumor.

I chased it down. The number traces to a single commercial aggregator that will not say what year it is from and shows a different figure, around 83, on another version of its own page. It is not in the FBI record. It is not in Tennessee's record. What is in the state record is a count of 110 non-consensual sex offenses in Johnson City in 2024, and a clearance rate of 26.36 percent. Fewer than three in ten cases closed. That is the number a person actually needs, and no tool offered it. The underlying records and links are documented in the Johnson City audit.

The obvious explanation is that local news died and the machines had nothing good to read. I believed that when I started. I spent 13 years running the websites for this region's newspapers, so I had a professional interest in believing it. Then I classified every wrong answer from three city audits, went and measured the sources myself, and found something narrower and stranger.

01 / The Ledger Thirty-nine wrong answers, sorted by why

Across the Kingsport, Johnson City, and Bristol audits I flagged 39 claims as wrong, stale, unsupported, or conspicuously missing. I went back through all of them and asked a different question than the one I asked the first time. Not what was wrong. Where the right answer was sitting, and whether a machine could get to it.

Two items out of 39. That is how often a paywalled local newspaper was the thing standing between a machine and the truth about this region. I expected that bar to be the tallest one on the chart. It is second from the bottom.

The tallest bar is stale commercial pages. A relocation site scrapes a Census figure once and leaves it up for four years. A crime-grade vendor runs a model on data that is two years old and prints a letter. A real estate aggregator quotes a 2020 home value in 2026. None of it is invented and all of it is free, fast, and beautifully formatted for a crawler. Accuracy and legibility are separate qualities, and only one of them is being rewarded.

02 / The Inversion The corrections were free. Nobody picked them up.

This is the part that broke my theory, so I am putting it early instead of burying it.

Three things got left out of all five tools in all three cities. The hospital monopoly. The Johnson City police failure. The Bristol landfill. Those are three of the most consequential facts surfaced in these audits, and every one of them was already published, in full, by somebody with money and a masthead.

The cleanest example is not any of those three. It is Eastman. The company's environmental file lives in EPA databases that are open, structured, downloadable, and built for exactly this kind of question. About 3.6 million pounds of reported toxic releases in 2024, roughly 95.7 percent of everything reported in Sullivan County. A Clean Air Act significant violation flag. The only sulfur dioxide nonattainment area in Tennessee. No paywall. No robots.txt. Five systems missed it anyway. The facility-level records and calculations are linked in the Kingsport audit.

There is research that explains the shape of this. LocalBench, published this spring, put more than 14,000 local questions covering 526 counties to the major models. The best one answered 56.8 percent of narrative questions correctly and under 15.5 percent of numerical ones. That is the same pattern I documented one city at a time. They can tell you what a town feels like. They cannot reliably tell you what it costs to live there.

And a Google research team found something worse than ignorance. Testing what happens when a model retrieves context that is real but insufficient, they measured one model's incorrect answers rising from 10.2 percent with no context at all to 66.1 percent with partial context. In that experiment, thin grounding was worse than none, because it gave the model just enough to answer incorrectly with confidence.

03 / The View From the Server What the local paper looked like to a machine, 2012 to 2025

I ran the websites for the Kingsport Times News and the Johnson City Press from 2012 until I stepped away from the director role in 2025. Four platforms in thirteen years. An in-house custom system, then Libercus around 2014, then TownNews BLOX around 2018 with a redesign in 2021, then mynews360 from Presteligence in 2025. Everybody writes about what happened to local news in the newsroom. Almost nobody writes about what it looked like from the server side, so here it is.

We never blocked the crawlers. Not once, on any platform. Our robots.txt allowed essentially the whole site and disallowed the admin paths, and that is still true today. The papers are not blocking the tested crawlers in robots.txt. That does not mean every article is equally discoverable or readable through a meter.

The losses came somewhere else, and they came at the seams. Every migration changed the URL structure and the naming convention. Every migration shrank the archive, because the practical scope was usually the last ten years and not the whole run. I argued against that every time, on the grounds that the older material was the part with historical value, and I lost that argument more than once. Every migration hurt us in search, and we felt it for months afterward. The one move that went smoothly was the BLOX redesign, and only because we stayed on the same platform.

There is a version of this story where a villain does something. There is no villain in it. Every one of those decisions was defensible on the day it was made, made by people under real budget pressure, with the archive quietly paying for it each time. That is what institutional memory loss actually looks like. It is not a fire. It is four reasonable meetings.

And the wider loss is measurable at the national level. Pew found that a quarter of web pages created between 2013 and 2023 are no longer reachable, and that 23 percent of news pages contain at least one broken link. The study did not isolate small local outlets, but there is no reason to treat their archives as exempt from the same failure modes.

04 / The Measurement What I found when I read the region's robots.txt files

On August 5, 2026 I pulled the crawler policy, the sitemaps, and the page-level structured data for every local news outlet and every city and county government in this region. Here is what is actually configured.

WJHL's robots.txt asks every major AI crawler I tested not to crawl. That includes training crawlers and live retrieval agents that could otherwise look up a correction in real time. The file also asks the Internet Archive not to crawl the site. WJHL was the single most-cited free source in all three audits, yet its published policy reduces direct access for current AI agents and requests exclusion from future archival capture.

WCYB's policy blocks most of them too, and asks the Internet Archive not to crawl its news, sports, and features sections. Tennessee Lookout's policy blocks them as well. That last one matters because Tennessee Lookout is one of the two outlets that documented the hospital monopoly. Part of the correction for the most consequential omission in the audits is therefore not directly retrievable by agents that respect that policy.

Then I checked whether any of those configurations were locally distinctive. They were not.

The observable crawler policies were inherited from corporate, nonprofit-network, and vendor templates rather than customized for Kingsport, Johnson City, or Bristol. I found no evidence of local customization. The files cannot establish who was consulted or why the defaults were chosen; they establish the configuration and the effect that landed here.

The government side is the same story in a different costume. The configuration audit covered six governments: Kingsport, Johnson City, Bristol, Tennessee, Bristol, Virginia, Sullivan County, and Washington County. Their meeting records run through outside portals and vendor systems. Kingsport's index sits behind a search form, with one board packet I pulled weighing 60.7 megabytes as a single PDF. Johnson City's portal disallows every crawler by default and carves out an exception for two search engines. Bristol and Washington County serve theirs as PDFs behind file-viewer links. None of the six published structured data for meetings.

To be fair about it, the documents themselves are fine. I checked 2026 records from five bodies: Kingsport, Johnson City, Bristol, Tennessee, Sullivan County, and Washington County. Bristol, Virginia was included in the configuration audit but not this document sample. The five sampled records were real text, not scanned images. You can search them. The problem is not that the record is unreadable. The problem is that nothing points to it, nothing describes it, and a 60 megabyte PDF behind a search form is a record that technically exists and practically does not.

05 / The Substitute Something already moved into the empty space

There is a national network of automated, local-news-shaped sites. Researchers commonly describe the category as "pink slime." The Metric Media Tennessee directory lists about 20 properties in the state, including two aimed at this region: the Kingsport Reporter and the Johnson City Times. Tow Center reporting has documented Metric Media's national scale, automated production, and political and corporate ties.

I went and read the Kingsport property. On August 5 its entire front page was Sullivan County jail bookings. Not most of it. All of it. Five people booked on property charges on July 30. Nine booked on warrant charges on July 31. At least 30 people booked at a county facility. The stories followed repeated templates, drew from booking records, appeared in batches, and were dated two or three days after the arrests.

The visible bylines were initials: A. C. Schmidt and T. E. Westbrook. I found no author pages or biographies behind them. The About page described a trusted source for community news and promised school-board coverage, local government, festivals, and human-interest features, but none of those categories appeared on the front page I measured. The site itself disclosed no owner, publisher, editor, staff list, street address, telephone number, or direct email address. Its Contact page provided a form.

Then I pulled its configuration the same way I pulled everyone else's, and that is the uncomfortable part.

Read that table again, because I had to. Among the three sources compared, the one easiest for a machine to discover and parse was an automated feed of who got booked into the county jail that week. It was free, indexed, structured, feed-enabled, and open to the tested live retrieval agents. The newspaper that has covered the city for more than a century did not block those agents but published no sitemap or article schema I could find. The station doing the most free reporting asked the tested agents not to crawl.

Sit with what that means for a minute. Somebody in Sullivan County gets booked on a Wednesday. By Friday their name can appear on a nationally operated website with a local name, formatted for machines, in the most machine-accessible source in this comparison. No local editor is disclosed. I could not establish that anyone attached to the site lives or works here. There is not even a published phone number.

That is the answer to the question in the title. When AI has no local news to read, it does not say so. It reads whatever is left. And what is left is getting better at being read than we are.

06 / The Bill What it costs a town to be described badly

It is easy to file this under trivia. A chatbot got the median home price wrong, so what. Here is the so what.

The vacuum did not fill only with bland tourism copy. In Kingsport it filled with the names of local residents on booking days, converted into templated pages that are easier for machines to discover than much of the reported record around them.

A family deciding whether to move here now asks a machine, and the machine tells them regional healthcare is a strength. That is the one thing in this whole search that could actually hurt them, and it is pointed exactly backward. A business owner wonders why the AI recommends a competitor, and the answer is a formatting decision nobody explained to them. A reporter looks for what happened here in 2016 and finds the archive was trimmed during a platform migration a decade ago.

And the state has already noticed the underlying problem, even if it has not connected it to this. Tennessee changed its public-notice law twice. When a legal notice must run in a newspaper, it must also appear on the newspaper's site and statewide repository. Public Chapter 793, effective July 1, 2024, added publication on a qualifying local news website if such a website exists. Public Chapter 500, effective July 1, 2025, requires that qualifying site to publish at least three times a week, keep an office in the county, run at least half original local editorial content, be separately owned from the newspaper, and charge no subscription to view. Read that requirement plainly. Where a qualifying independent site exists, the legislature decided that publication through the newspaper and its own website alone is not enough.

07 / The Fix Boring, cheap, and available today

Several of the fixes in this piece are configuration and publishing work rather than new policy or a new grant. They require somebody to make them a priority.

For the papers. Publish a sitemap and a news sitemap. Put article schema on article pages. Keep the archive whole through the next migration and redirect the old URLs. Leave the free content free and mark it that way, so a machine can tell the difference between what is metered and what is not. Six Rivers has no AI licensing agreement I could find as of this snapshot, and the News/Media Alliance now offers an opt-in licensing program for small and midsize publishers. That is at least a conversation worth having.

For the stations. The matching files show a network-level policy that the local newsroom inherits. There is a real argument for blocking training crawlers and a weaker one for blocking live retrieval agents, because doing so reduces the chance that your reporting appears in a current answer. Asking the Internet Archive not to crawl also increases the risk that part of the region's record will not be preserved there, for a benefit that is not obvious to me.

For the cities and counties. Put the agenda on the page as text, not only inside a packet. Split the 60 megabyte PDF. Let the meeting index be crawlable instead of hiding behind a search form. This can begin as a vendor-configuration conversation; the actual effort and cost will depend on the contract and platform.

For everybody else. Do not let a machine have the last word on anything that matters. That is not a slogan, it is what the numbers say. Under 15.5 percent correct on local numerical questions is not a tool you trust with your mortgage or your hospital.

08 / Method How this was measured

The 39 classified errors come from the Kingsport, Johnson City, and Bristol audits published on this site in June 2026. Each audit ran five AI research systems against the same question, then verified every returned figure against a primary source. For this piece I went back through every claim I had tagged as wrong, stale, unsupported, or missing, and assigned each one the obstacle that would have had to come down first for a system to reach the right answer.

The crawler policy, sitemap, and structured data findings were pulled directly from each site on August 5, 2026, and the file-matching between sister sites was confirmed by hashing. That is a snapshot of one day. Any of it can change without notice, and I would rather it did.

Three limits are worth stating. A robots.txt directive is a request and not a wall, so where a site tells the Internet Archive to stay away, treat that as stated intent rather than a measured outcome. Nobody has published a rationale for these configurations, so the files establish effect, not motive or consultation. And although Metric Media's directory establishes the Kingsport Reporter's network affiliation, the public pages I reviewed did not establish whether any contributor is physically based in Northeast Tennessee.

Questions people actually ask

Snapshot: August 5, 2026. Configurations move. If you find one of these findings out of date, that is not a strike against the method. That is the method working.